Osm2pgsql Manual

Introduction

Osm2pgsql is used to import OSM data into a PostgreSQL/PostGIS database for rendering into maps and many other uses. Usually it is only part of a toolchain, for instance other software is needed for the actual rendering (i.e. turning the data into a map) or delivery of the maps to the user etc.

Osm2pgsql is a fairly complex piece of software and it interacts with the database system and other pieces of the toolchain in complex ways. It will take a while until you get some experience with it. It is strongly recommended that you try out osm2pgsql with small extracts of the OSM data, for instance the data for your city. Do not start with importing data for the whole planet, it is easy to get something wrong and it will break after processing the data for hours or days. You can save yourself a lot of trouble doing some trial runs starting with small extracts and working your way up, observing memory and disk usage along the way.

It helps to be familiar with the PostgreSQL database system and the PostGIS extension to it as well as the SQL database query language. Some knowledge of the Lua language is also useful.

This manual always documents the current version of osm2pgsql. If same information is only valid in certain versions, the section will have a note like this: Version >= 1.4.0

Sections labelled Experimental describe new features which might change at any point without notice. We encourage you to experiment with them and report how they work for you, but don’t rely on then for production use.

It is recommended that you always use the newest released version of osm2pgsql. Earlier versions sometimes contain bugs that have long since been fixed.

System Requirements

Operating System

Osm2pgsql works on Linux, Windows, and macOS.

Osm2pgsql is developed on Linux and most of the developers don’t have experience with running it on anything but Linux, so it probably functions best on that platform. This documentation is also somewhat geared towards Linux users. That being said, we strive to make the software work on Windows and macOS as well. Please report any problems you might have.

Only 64bit systems are supported.

Read the installation instructions for details on how to install osm2pgsql on your system.

Main Memory

Memory requirements for your system will vary widely depending on the size of your input file. Is it city-sized extract or the whole planet? As a rule of thumb you need at least as much main memory as the PBF file with the OSM data is large. So for a planet file you currently need at least 64 GB RAM. Osm2pgsql will not work with less than 2 GB RAM.

More memory can be used as cache and speed up processing a lot.

Disk

You are strongly encouraged to use a SSD (NVMe if possible) for your database. This will be much faster than traditional spinning hard disks.

Requirements for your system will vary widely depending on

• the amount of data you want to store (city-sized extract or the whole planet?)
• whether or not you want to update the data regularly

Database Software

You need the PostgreSQL database system with the PostGIS extension installed.

Osm2pgsql aims to support all PostgreSQL and PostGIS versions that are currently supported by their respective maintainers. Currently PostgreSQL versions 9.6 and above and PostGIS versions 2.2 and above are supported. (Earlier versions might work but are not tested any more.) PostgreSQL version 11 or above and PostGIS version 2.4 or above is recommended.

In some cases older versions of osm2pgsql have problems with newer database software versions. You should always run the newest released osm2pgsql version which should ensure you get any fixes we have done to make osm2pgsql work well with any PostgreSQL version.

Osm2pgsql does not work with database systems other than PostgreSQL. The PostgreSQL/PostGIS combination is unique in its capabilities and we are using some of their special features. We are not planning to support any other database systems. There are some database systems out there which claim compatibility with PostgreSQL, they might work or they might not work. Please tell us if you have any experience with them.

Lua Scripting Language

Some parts of osm2pgsql require the use of the Lua scripting language. It is highly recommended that you use a version of osm2pgsql with Lua support enabled, because it gives you a much more flexible configuration. If available osm2pgsql can also be compiled using the Lua JIT (just in time) compiler.

Osm2pgsql can also be compiled without Lua support. In that case the pgsql and gazetteer outputs are the only ones available.

Security Note: Lua scripts can do basically anything on your computer that the user running osm2pgsql is allowed to do, such as reading and writing files, opening network connections etc. This makes the Lua config files really versatile and allows lots of great functionality. But you should never use a Lua configuration file from an unknown source without checking it.

Preparing the Database

Before you can import any OSM data into a database, you need a database.

Installing PostgreSQL/PostGIS

You need the PostgreSQL database software and the PostGIS plugin for the database. Please read the respective documentation on how to install it. On Linux this can almost always be done with the package manager of your distribution.

• PostgreSQL download page
• PostGIS download page

Creating a Database

To create a database that can be used by osm2pgsql follow these steps:

1. Create a database user that osm2pgsql will use. This user doesn’t need any special rights. We’ll use osmuser here.
2. Create a database that osm2pgsql will use belonging to the user you just created. We’ll use osm as a name here.
3. Enable the postgis and hstore extensions in the newly created database.

On a typical Linux system you’ll have a system user called postgres which has admin privileges for the database system. We’ll use that user for these steps.

Here are typical commands used:

sudo -u postgres createuser osmuser
sudo -u postgres createdb --encoding=UTF8 --owner=osmuser osm
sudo -u postgres psql osm --command='CREATE EXTENSION postgis;'
sudo -u postgres psql osm --command='CREATE EXTENSION hstore;'

Security Considerations

Osm2pgsql does not need any special database rights, it doesn’t need superuser status and doesn’t need to create databases or roles. You should create a database user for the specific use by osm2pgsql which should not have any special PostgreSQL privileges.

This manual assumes that you have ‘peer’ authentication configured when using a local database. That means a local user may connect to the database with its own username without needing to type a password. This authentication method is the default on Debian/Ubuntu-based distributions. Never configure ‘trust’ authentication where all local users can connect as any user to the database.

Any osm2pgsql setup will need a database to work on. You should create this as PostgreSQL superuser and change ownership (with --owner option on the createdb command or an OWNER= clause on the CREATE DATABASE SQL command) to the user osm2pgsql is using. This way the database user that osm2pgsql uses doesn’t need database creation privileges.

Typical osm2pgsql setups need the postgis and hstore extensions to be enabled in the database. To install these you need superuser privileges in the database. Enable them (with CREATE EXTENSION) as PostgreSQL superuser on the database that osm2pgsql is using before you run osm2pgsql.

Of course osm2pgsql needs to be able to create tables and write to the database. Usually it can do this as owner of the database created for it. Using the data, on the other hand, doesn’t need any of those rights. So the map rendering software you are using, for instance, usually only needs to read the data. It is recommended that you run these as a different database user, distinct from the database user osm2pgsql is using and only give that user SELECT rights (with the GRANT command).

Version >= 1.4.0 If you are using a security scheme based on database schemas in your database you can use the --middle-schema and --output-pgsql-schema options and the schema table option in the flex output, respectively, to tell osm2pgsql to load data into specific schemas. You have to create those schemas and give them the correct rights before running osm2pgsql.

Before PostgreSQL 15 all database users could add objects (such as tables and indexes) to the public schema by default. Since PostgreSQL 15, by default, only the owner of the database is allowed to do this. If osm2pgsql is using a database user that’s not the owner of the database, you have to grant this user CREATE rights on the public schema of the database, or you have to configure osm2pgsql to use a different schema as described above. See the PostgreSQL manual for the details.

Encoding

OpenStreetMap data is from all around the world, it always uses UTF-8 encoding. osm2pgsql will write the data as is into the database, so it has to be in UTF-8 encoding, too.

On any modern system the default encoding for new databases should be UTF-8, but to make sure, you can use the -E UTF8 or --encoding=UTF8 options when creating the database for osm2pgsql with createdb.

Tuning the PostgreSQL Server

Usual installs of the PostgreSQL server come with a default configuration that is not well tuned for large databases. You should change these settings in postgresql.conf and restart PostgreSQL before running osm2pgsql, otherwise your system will be much slower than necessary.

The following settings are geared towards a system with 64GB RAM and a fast SSD. The values in the second column are suggestions to provide a good starting point for a typical setup, you might have to adjust them for your use case. The value in the third column is the default set by PostgreSQL 13.

Here are the proposed settings for copy-and-paste into a config file:

shared_buffers = 1GB
work_mem = 50MB
maintenance_work_mem = 10GB
autovacuum_work_mem = 2GB
wal_level = minimal
checkpoint_timeout = 60min
max_wal_size = 10GB
checkpoint_completion_target = 0.9
max_wal_senders = 0
random_page_cost = 1.0

Increasing values for max_wal_size and checkpoint_timeout means that PostgreSQL needs to run checkpoints less often but it can require additional space on your disk and increases time required for Point in Time Recovery (PITR) restores. Monitor the Postgres logs for warnings indicating checkpoints are occurring too frequently: HINT: Consider increasing the configuration parameter "max_wal_size".

Autovacuum must not be switched off because it ensures that the tables are frequently analysed. If your machine has very little memory, you might consider setting autovacuum_max_workers = 1 and reduce autovacuum_work_mem even further. This will reduce the amount of memory that autovacuum takes away from the import process.

For additional details see the Server Configuration chapter and Populating a Database in the Performance Tips chapter in the PostgreSQL documentation.

Expert Tuning

The suggestions in this section are potentially dangerous and are not suitable for all environments. These settings can cause crashes and/or corruption. Corruption in a Postgres instance can lead to a “bricked” instance affecting all databases in the instance.

Additional information is on the PostgreSQL wiki: Tuning Your PostgreSQL Server. The section titled synchronous_commit contains important information to the synchronous_commit and fsync settings.

Number of Connections

Osm2pgsql will open multiple connections to the database to speed up the import. The number of connections will depend on the number of tables that are configured and the number of threads used. This can easily exceed the max_connections limit defined in the PostgreSQL config.

(Currently there are about 3 + number of tables connections used on import and 3 + (1 + number of threads) * number of tables connections used on update. But this might change from version to version of osm2pgsql.)

If you are hit by this, your options are to

• increase the max_connections settings in your database configuration, or
• reduce the number of threads with the --number-processes=THREADS command line option of osm2pgsql, or

Using a Template Database

Databases are actually created in PostgreSQL by copying a template database. If you don’t specify a database to copy from, the template1 database is used. PostgreSQL allows you to change template1 or to create additional template databases. If you create a lot of databases for use with osm2pgsql, you can do initializations like CREATE EXTENSTION postgis; once in a template database and they will be available in any database you create from them.

See the PostgreSQL manual for details.

Geometry Processing

An important part of what osm2pgsql does is creating geometries from OSM data. In OSM only nodes have a location, ways get their geometry from member nodes and relations get their geometry from member nodes and ways. Osm2pgsql assembles all the data from the related objects into valid geometries.

Geometry Types

The geometry types supported by PostGIS are from the Simple Features defined by the OpenGIS Consortium (OGC). Osm2pgsql creates geometries of these types from OSM data.

Single vs. Multi Geometries

Generally osm2pgsql will create the simplest geometry it can. Nodes will turn into Points, ways into LineStrings or Polygons. A multipolygon relation can be turned into a Polygon or MultiPolygon, depending on whether it has one or more outer rings. Similarly, a route relation can be turned into a LineString if the route is connected from start to finish or a MultiLineString if it is unconnected or there are places, where the route splits up.

In some cases osm2pgsql will split up Multi* geometries into simple geometries and add each one in its own database row. This can make rendering faster, because the renderer can deal with several smaller geometries instead of having to handle one large geometry. But, depending on what you are doing with the data, it can also lead to problems. This blogpost has some deeper discussion of this issue. See the flex and pgsql output chapters for details on how to configure this. It will also mean that your id columns are not unique, because there are now multiple rows created from the same OSM object. See the Primary Keys and Unique IDs section for an option how to work around this.

Version >= 1.7.0 When using the flex output, you can decide yourself what geometries to create using the as_point(), as_linestring(), as_polygon(), as_multipoint(), as_multilinestring(), as_multipolygon(), and as_geometrycollection() functions. See the Flex Output chapter for details.

Geometry Validity

Point geometries are always valid (as long as the coordinates are inside the correct range). LineString geometries are valid if they have at least two distinct points. Osm2pgsql will collapse consecutive indentical points in a linestring into a single point. Note that a line crossing itself is still valid and not a problem.

Validity is more complex for Polygon and MultiPolygon geometries: There are multiple ways how such a geometry can be invalid. For instance, if the boundary of a polygon is drawn in a figure eight, the result will not be a valid polygon. The database will happily store such invalid polygons, but this can lead to problems later, when you try to draw them or do calculations (such as the area) based on the invalid geometry.

You can use the Areas view of the OpenStreetMap inspector to help diagnose problems with multipolygons.

Osm2pgsql makes sure that there are no invalid geometries in the database, either by not importing them in the first place or by using an ST_IsValid() check after import. You’ll either get NULL in your geometry columns instead or the row is not inserted at all (depending on config).

The transform() function in Lua config files projects a geometry into a different SRS. In special cases this can make the geometry invalid, for instance when two distinct points are projected onto the same point.

Processing of Nodes

Node geometries are always converted into Point geometries.

Processing of Ways

Depending on the tags, OSM ways model either a LineString or a Polygon, or both! A way tagged with highway=primary is usually a linear feature, a way tagged landuse=farmland is usually a polygon feature. If a way with polygon type tags is not closed, the geometry is invalid, this is an error and the object is ignored. For some tags, like man_made=pier non-closed ways are linear features and closed ways are polygon features.

If a mapper wants to override how a way should be interpreted, they can use the area tag: The tag area=yes turns a normally linear feature into a polygon feature, for instance it turns a pedestrian street (highway=pedestrian) into a pedestrian area. The tag area=no turns a polygon feature into a linear feature.

There is no definite list which tags indicate a linear or polygon feature. Osm2pgsql lets the user decide. It depends on your chosen output (see next chapter) how to configure this. The example config files have lists that should cover most of the commonly used tags, but you might have to extend the lists if you are using more unusual tags.

Osm2pgsql can split up long LineStrings created from ways into smaller segments. This can make rendering of tiles faster, because smaller geometries need to be retrieved from the database when rendering a specific tile. The pgsql output always splits up long LineStrings, in latlong projection, lines will not be longer than 1°, in Web Mercator lines will not be longer than 100,000 units (about 100,000 meters at the equator). The flex output will only split LineStrings if the split_at transformation parameter is used, see the Geometry transformations section in the flex output chapter for details. See also the Single vs. Multi Geometries section above.

Version >= 1.7.0 When using the flex output, you can decide yourself what geometries to create from a way using the as_linestring(), or as_polygon() functions. See the Flex Output chapter for details.

Processing of Relations

Relations come in many variations and they can be used for all sorts of geometries. Usually it depends on the type tag of a relation what kind of geometry it should have:

The osm2pgsql flex output can create a (Multi)Polygon or (Multi)LineString geometry from any relation, other geometries are currently not supported. See the Geometry transformations section in the flex output chapter for details.

Version >= 1.7.0 When using the flex output, you can decide yourself what geometries to create from a way using the as_multipoint(), as_multilinestring(), or as_multipolygon() functions. Also supported now is the as_geometrycollection() function which creates GeometryCollection from all member nodes and ways of a relation (relation members are ignored).

If you are using the old “C transform” of the pgsql output, the geometry types for relations of type multipolygon, boundary, and route are hardcoded. If you are using the “Lua transform” you can configure them.

Note that osm2pgsql will ignore the roles (inner and outer) on multipolygon and boundary relations when assembling multipolygons, because the roles are sometimes wrong or missing.

Handling of Incomplete OSM Data

Sometimes you will feed incomplete OSM data to osm2pgsql. Most often this will happen when you use geographical extracts, either data you downloaded from somewhere or created yourself. Incomplete data, in this case, means that some member nodes of ways or members of relations are missing. Often this can not be avoided when creating an extract, you just have to put that cut somewhere.

When osm2pgsql encounters incomplete OSM data it will still try to do its best to use it. Ways missing some nodes will be shortened to the available part. Multipolygons might be missing some parts. In most cases this will work well (if your extract has been created with a sufficiently large buffer), but sometimes this will lead to wrong results. In the worst case, if a complete outer ring of a multipolygon is missing, the multipolygon will appear “inverted”, with outer and inner rings switching their roles.

Unfortunately there isn’t much that osm2pgsql (or anybody) can do to improve this, this is just the nature of OSM data.

Projections

Osm2pgsql can create geometries in many projections. If you are using the pgsql output, the projection can be chosen with command line options. When using the flex output, the projections are specified in the Lua style file. The default is always “Web Mercator”.

Version >= 1.7.0 When using the flex output, osm2pgsql will usually magically transform any geometry you are writing into a database table into the projection you defined your tables with. But you can use the transform() function on the geometry to force a certain transformation. This is useful, for instance, if you want to calculate the area of a polygon in a specific projection. See the Flex Output chapter for details.

	Latlong (WGS84) EPSG:4326 The original latitude & longitude coordinates from OpenStreetMap in the WGS84 coordinate reference system. This is typically chosen if you want to do some kind of analytics on the data or reproject it later.
	Web Mercator EPSG:3857 This is the projection used most often for tiled web maps. It is the default in osm2pgsql. Data beyond about 85° North and South will be cut off, because it can not be represented in this projection.
	Other Projections If osm2pgsql was compiled with support for the PROJ library, it supports all projections supported by that library. Version >= 1.4.0 Call osm2pgsql --version to see whether your binary was compiled with PROJ and with which version.

Note that mapping styles often depend on the projection used. Most mapping style configurations will enable or disable certain rendering styles depending on the map scale or zoom level. But a meaningful scale will depend on the projection. Most styles you encounter are probably made for Web Mercator and will need changes to get nice maps in other projections.

Running osm2pgsql

Basic command line operation

Osm2pgsql can work in one of two ways: Import only or Import and Update.

If you are new to osm2pgsql we recommend you try the import only approach first and use a small extract of the OSM data.

Import Only

OSM data is imported once into the database and will not change afterwards. If you want to update the data, you have to delete it and do a full re-import. This approach is used when updates aren’t needed or only happen rarely. It is also sometimes the only option, for instance when you are using extracts for which change files are not available. This is also a possible approach if disk space is tight, because you don’t need to store data needed only for updates.

In this mode osm2pgsql is used with the -c, --create command line option. This is also the default mode. The following command lines are equivalent:

osm2pgsql -c OSMFILE
osm2pgsql --create OSMFILE
osm2pgsql OSMFILE

In case the system doesn’t have much main memory, you can add the -s, --slim and --drop options. In this case less main memory is used, instead more data is stored in the database. This makes the import much slower, though. See the chapter Middle for details on these options.

Import and Update

In this approach OSM data is imported once and afterwards updated more or less regularly from OSM change files. OSM offers minutely, hourly, and daily change files. This mode is used when regular updates are desired and change files are available.

In this mode osm2pgsql is used the first time with the -c, --create command line option (this is also the default mode) and, additionally the -s, --slim options. The following command lines are equivalent:

osm2pgsql -c -s OSMFILE
osm2pgsql --create --slim OSMFILE
osm2pgsql --slim OSMFILE

For the update runs the -a, --append and -s, --slim options must be used. The following command lines are equivalent:

osm2pgsql -a -s OSMFILE
osm2pgsql --append --slim OSMFILE

The OSMFILE in this case will usually be an OSM change file (with suffix .osc or .osc.gz).

You can not use replication diffs downloaded from planet.osm.org directly with osm2pgsql, see Updating an Existing Database for details.

This approach needs more disk space for your database than the “Import Only” approach, because all the information necessary for the updates must be stored somewhere.

Usually you will need to use at least the -C, --cache and --flat-nodes command line options when doing imports and updates. See the chapter Middle for details.

Getting Help or Version

To get help or the program version call osm2pgsql with one of the following options:

Logging

Osm2pgsql will write information about what it is doing to the console (to STDERR). If you’d rather want this information in a file, use the output redirection in your shell (osm2pgsql ... 2>osm2pgsql.log).

Several command line options allow you to change what will be logged and how:

Database Connection

In create and append mode you have to tell osm2pgsql which database to access. If left unset, it will attempt to connect to the default database (usually the username) using a unix domain socket. Most usage only requires setting -d, --database.

You can also use libpq environment variables to set connection parameters. For a full list of available parameters, please consult the PostgreSQL documentation.

When you need a password for your database connection and want to run osm2pgsql inside scripts, then use a pgpass file with appropriate permissions to store the password. All other methods of giving the password to osm2pgsql are inherently insecure. You can either put a .pgpass file in the user’s home directory or supply its location through the PGPASSFILE environment variable.

Version >= 1.4.0 Instead of specifying a database name with the -d, --database option you can also specify a connection string in the form of a keyword/value connection string (something like host=localhost port=5432 dbname=mydb) or a URI (postgresql://[user[:password]@][netloc][:port][,...][/dbname][?param1=value1&...]) See the PostgreSQL documentation for details.

Processing the OSM Data

Osm2pgsql processing can be separated into multiple steps:

1. The Input reads the OSM data from the OSM file.
2. The Middle stores all objects and keeps track of relationships between objects.
3. The Output transforms the data and loads it into the database.

This is, of course, a somewhat simplified view, but it is enough to understand the operation of the program. The following sections will describe each of the steps in more detail.

The Input

Depending on the operational mode your input files are either

• OSM data files (in create mode), or
• OSM change files (in append mode)

Usually osm2pgsql will autodetect the file format, but see the -r, --input-reader option below. Osm2pgsql can not work with OSM history files.

OSM data files are almost always sorted, first nodes in order of their ids, then ways in order of their ids, then relations in order of their ids. The planet files, change files, and usual extracts all follow this convention.

Osm2pgsql can only read OSM files ordered in this way. This allows some optimizations in the code which speed up the normal processing.

Older versions of osm2pgsql will sometimes work or appear to work with unsorted data, but it is not recommended to rely on this, because the processed data might be corrupted. Versions from 1.3.0 warn when you are using unsorted data. From version 1.4.0 on, only sorted OSM files are allowed.

See the Appendix A for more information on how to get and prepare OSM data for use with osm2pgsql.

Use of the -b, --bbox option is not recommended; it is a crude way of limiting the amount of data loaded into the database and it will not always do what you expect it to do, especially at the boundaries of the bounding box. If you use it, choose a bounding box that is larger than your actual area of interest. A better option is to create an extract of the data before using osm2pgsql. See Appendix A for options.

Working with Multiple Input Files

Usually you are using osm2pgsql with a single input file.

Do not use osm2pgsql with more than one input file in versions before 1.4.0. It might work or it might not.

Version >= 1.4.0 Osm2pgsql can read multiple input files at once, merging the data from the input files ignoring any duplicate data. For this to work the input files must all have their data from the same point in time. You can use this to import two or more geographical extracts into the same database. If the extracts are from different points in time and contain different versions of the same object, this will fail!

Do not use multiple change files as input in append mode, merge and simplify them first.

The Middle

The middle keeps track of all OSM objects read by osm2pgsql and the relationships between those objects. It knows, for instance, which ways are used by which nodes, or which members a relation has. It also keeps track of all node locations. This information is necessary to build way geometries from way nodes and relation geometries from members and it is necessary when updating data, because OSM change files only contain changes objects themselves and not all the related objects needed for creating an objects geometry.

More details are in its own chapter.

The Outputs

Osm2pgsql imports OSM data into a PostgreSQL database. How it does this is governed by the output (sometimes called a backend). Several outputs for different use cases are available:

The flex Output

Available as experimental from version >= 1.3.0, stable from version >= 1.5.0

This is the most modern and most flexible output option. If you are starting a new project, use this output. Many future improvements to osm2pgsql will only be available in this output.

Unlike all the other output options there is almost no limit to how the OSM data can be imported into the database. You can decide which OSM objects should be written to which columns in which database tables. And you can define any transformations to the data you need, for instance to unify a complex tagging schema into a simpler schema if that’s enough for your use case.

This output is described in detail in its own chapter.

The pgsql Output

The pgsql output is the original output and the one that most people who have been using osm2pgsql know about. Many tutorials you find on the Internet only describe this output. It is quite limited in how the data can be written to the database, but many setups still use it.

This output comes in two “flavours”: With the original “C transformation” and with the somewhat newer “Lua transformation” which allows some changes to the data before it is imported.

This output is described in detail in its own chapter.

The gazetteer Output

The gazetteer output is a specialized output used for Nominatim only. There is no information in this manual about its use, see the Nominatim documentation for details.

Version >= 1.8.0 The gazetteer output is deprecated.

The multi Output

Version < 1.5.0 The multi output is deprecated.

Version >= 1.5.0 The multi output was removed.

The null Output

The null output doesn’t write the data anywhere. It is used for testing and benchmarking, not for normal operation. If --slim is used with the null output, the middle tables are still generated.

Here are the command line options pertaining to the outputs (see later chapters for command line options only used for specific outputs):

Environment Variables

Osm2pgsql itself doesn’t interpret any environment variables, but several of the libraries it uses do.

• The PostgreSQL database access library (libpq) understands many environment variables.
• The libosmium library has some internal settings you can change.
• If osm2pgsql was compiled with the PROJ library, you can use various environment variables to control the behaviour of the library.

You can access environment variables from Lua config scripts with os.getenv("VAR").

The Flex Output

Version 1.3.0 to 1.4.2 The flex output appeared first in version 1.3.0 of osm2pgsql. It was marked as experimental until version 1.4.2

The flex output, as the name suggests, allows for a flexible configuration that tells osm2pgsql what OSM data to store in your database and exactly where and how. It is configured through a Lua file which

• defines the structure of the output tables and
• defines functions to map the OSM data to the database data format

Use the -S, --style=FILE option to specify the name of the Lua file along with the -O flex or --output=flex option to specify the use of the flex output.

Unlike the pgsql output, the flex output doesn’t use command line options for configuration, but the Lua config file only.

The flex style file is a Lua script. You can use all the power of the Lua language. It gives you a lot of freedom to write complex preprocessing scripts that might even use external libraries to extend the capabilities of osm2pgsql. But it also means that the scripts may mess with any part of your system when written badly. Only run Lua scripts from trusted sources!

This description assumes that you are somewhat familiar with the Lua language, but it is pretty easy to pick up the basics and you can use the example config files in the flex-config directory which contain lots of comments to get you started.

All configuration is done through the osm2pgsql global object in Lua. It has the following fields and functions:

Osm2pgsql also provides some additional functions in the Lua helper library described in Appendix B.

Defining a Table

You have to define one or more database tables where your data should end up. This is done with the osm2pgsql.define_table() function or, more commonly, with one of the slightly more convenient functions osm2pgsql.define_(node|way|relation|area)_table(). In create mode, osm2pgsql will create those tables for you in the database.

Basic Table Definition

The simple way to define a table looks like this:

osm2pgsql.define_(node|way|relation|area)_table(NAME, COLUMNS[, OPTIONS])

Here NAME is the name of the table, COLUMNS is a list of Lua tables describing the columns as documented below. OPTIONS is a Lua table with options for the table as a whole. An example might be more helpful:

local restaurants = osm2pgsql.define_node_table('restaurants', {
    { column = 'name', type = 'text' },
    { column = 'tags', type = 'jsonb' },
    { column = 'geom', type = 'point' }
})

In this case we are creating a table that is intended to be filled with OSM nodes, the table called restaurants will be created in the database with four columns: A name column of type text (which will presumably later be filled with the name of the restaurant), a column called tags with type jsonb (which will presumable be filled later with all tags of a node), a column called geom which will contain the Point geometry of the node and an Id column called node_id.

Each table is either a node table, way table, relation table, or area table. This means that the data for that table comes primarily from a node, way, relation, or area, respectively. Osm2pgsql makes sure that the OSM object id will be stored in the table so that later updates to those OSM objects (or deletions) will be properly reflected in the tables. Area tables are special, they can contain data derived from ways and from (multipolygon) relations.

Advanced Table Definition

Sometimes the define_(node|way|relation|area)_table() functions are a bit too restrictive, for instance if you want more control over the type and naming of the Id column(s). In this case you can use the function osm2pgsql.define_table().

Here are the available OPTIONS for the osm2pgsql.define_table(OPTIONS) function. You can use the same options on the define_(node|way|relation|area)_table() functions, except the name and columns options.

All the osm2pgsql.define*table() functions return a database table Lua object. You can call the following functions on it:

Id Handling

Id columns allows osm2pgsql to track which database table entries have been generated by which objects. When objects change or are deleted, osm2pgsql uses the ids to update or remove all existing rows from the database with those ids.

If you are using the osm2pgsql.define_(node|way|relation|area)_table() convenience functions, osm2pgsql will automatically create an id column named (node|way|relation|area)_id, respectively.

If you want more control over the id column(s), use the osm2pgsql.define_table() function. You can then use the ids option to define how the id column(s) should look like. To generate the same “restaurants” table described above, the define_table() call will look like this:

local restaurants = osm2pgsql.define_table({
    name = 'restaurants',
    ids = { type = 'node', id_column = 'node_id' },
    columns = {
        { column = 'name', type = 'text' },
        { column = 'tags', type = 'jsonb' },
        { column = 'geom', type = 'point' }
}})

If you don’t have an ids field, the table is generated without Id. Tables without Id can not be updated, so you can fill them in create mode, but they will not work in append mode.

The following values are allowed for the type of the ids field:

The any type is for tables that can store any type of OSM object (nodes, ways, or relations). There are two ways the id can be stored:

1. If you have a type_column setting in your ids field, it will store the type of the object in an additional char(1) column as N, W, or R.
2. If you don’t have a type_column, the id of nodes stays the same (so they are positive), way ids will be stored as negative numbers and relation ids are stored as negative numbers and 1e17 is subtracted from them. This results in distinct ranges for all ids so they can be kept apart. (This is the same transformation that the Imposm program uses.)

Unique Ids

It is often desirable to have a unique PRIMARY KEY on database tables. Many programs need this.

There seems to be a natural unique key, the OSM node, way, or relation ID the data came from. But there is a problem with that: Depending on the configuration, osm2pgsql sometimes adds several rows to the output tables with the same OSM ID. This typically happens when long linestrings are split into shorter pieces or multipolygons are split into their constituent polygons, but it can also happen if your Lua configuration adds two rows to the same table.

If you need unique keys on your database tables there are two options: Using those natural keys and making sure that you don’t have duplicate entries. Or adding an additional ID column. The latter is easier to do and will work in all cases, but it adds some overhead.

Using Natural Keys for Unique Ids

To use OSM IDs as primary keys, you have to make sure that

• you only ever add a single row per OSM object to an output table, i.e. do not call add_row multiple times on the same table for the same OSM object.
• osm2pgsql doesn’t split long linestrings into smaller ones or multipolygons into polygons. So you can not use the split_at option on the geometry transformation.
• Version == 1.3.0 osm2pgsql doesn’t split multipolygons into polygons. So you have to set multi = true on all area geometry transformations.

You probably also want an index on the ID column. If you are running in slim mode, osm2pgsql will create that index for you. But in non-slim mode you have to do this yourself with CREATE UNIQUE INDEX. You can also use ALTER TABLE to make the column an “official” primary key column.

Using an Additional ID Column

PostgreSQL has the somewhat magic “serial” data type. If you use that datatype in a column definition, PostgreSQL will add an integer column to the table and automatically fill it with an autoincrementing value.

In the flex config you can add such a column to your tables with something like this:

...
{ column = 'id', sql_type = 'serial', create_only = true },
...

The create_only tells osm2pgsql that it should create this column but not try to fill it when adding rows (because PostgreSQL does it for us).

You probably also want an index on this column. After the first import of your data using osm2pgsql, use CREATE UNIQUE INDEX to create one. You can also use ALTER TABLE to make the column an “official” primary key column.

Since PostgreSQL 10 you can use the GENERATED ... AS IDENTITY clause instead of the SERIAL type which does something very similar, although using anything but a proper PostgreSQL type here is not officially supported.

Defining Columns

In the table definitions the columns are specified as a list of Lua tables with the following keys:

The type field describes the type of the column from the point of view of osm2pgsql, the sql_type describes the type of the column from the point of view of the PostgreSQL database. Usually they are either the same or the SQL type is derived directly from the type according to the following table. But it is possible to set them individually for special cases.

Version < 1.5.0 The sql_type field did not exist in earlier versions. Use the type field instead. If you are upgrading from a version < 1.5.0 to 1.5.0 or above, it is usually enough to set the sql_type to the same value that the type field had before for every type field where you get an error message from osm2pgsql.

The SRID for the geometry SQL types comes from the projection parameter.

For special cases you usually want to keep the type field unset and only set the sql_type field. So if you want to have an array of integer column, for instance, set only the sql_type to int[]. The type field will default to text which means you have to create the text representation of your array in the form that PostgreSQL expects it.

For details on how the data is converted depending on the type, see the section on Type Conversions.

The content of the sql_type field is not checked by osm2pgsql, it is passed on to PostgreSQL as is. Do not set this to anything else but a valid PostgreSQL type.

Defining Geometry Columns

Most tables will have a geometry column. The types of the geometry column possible depend on the type of the input data. For node tables you are pretty much restricted to point geometries, but there is a variety of options for relation tables for instance.

Version < 1.7.0 Only zero or one geometry columns are supported.

Version >= 1.7.0 You can have any number of geometry columns if you are using the insert() command on a table (and not add_row(), see below).

An index will only be built automatically for the first (or only) geometry column you define. The table will be clustered by the first (or only) geometry column (unless disabled).

The supported geometry types are:

By default geometry columns will be created in web mercator (EPSG 3857). To change this, set the projection parameter of the column to the EPSG code you want (or one of the strings latlon(g), WGS84, or merc(ator), case is ignored).

There is one special geometry column type called area. It can be used in addition to a polygon or multipolygon column. Unlike the normal geometry column types, the resulting database type will not be a geometry type, but real. It will be filled automatically with the area of the geometry. The area will be calculated in web mercator, or you can set the projection parameter of the column to 4326 to calculate it with WGS84 coordinates. Other projections are currently not supported.

Version >= 1.7.0 If you are using the insert() command on a table (and not add_row(), see below) you can use the area() function on any geometry in any projection to calculate the area. What you get back is a real number that you can put into any normal real column. The special case of the area column type is not needed any more.

Defining Indexes

Osm2pgsql will always create indexes on the id column(s) of all tables if the database is updateable (i.e. in slim mode), because it needs those indexes to update the database. You can not control those indexes with the settings described in this section.

Version >= 1.8.0 Indexes can only be defined in version 1.8.0 and above, before that there was always a GIST index created on the first (or only) geometry column of any table.

To define indexes, set the indexes field of the table definition to an array of Lua tables. If the array is empty, no indexes are created for this table (except possibly an index on the id column(s)). If there is no indexes field (or if it is set to nil) a GIST index will be created on the first (or only) geometry column of this table. This is the same behavior as before version 1.8.0.

The following fields can be set in an index definition. You have to set at least the method and either column or expression.

If you need an index that can not be expressed with these definitions, you have to create it yourself using the SQL CREATE INDEX command after osm2pgsql has finished its job.

Processing Callbacks

You are expected to define one or more of the following functions:

They all have a single argument of type table (here called object) and no return value. If you are not interested in all object types, you do not have to supply all the functions.

These functions are called for each new or modified OSM object in the input file. No function is called for deleted objects, osm2pgsql will automatically delete all data in your database tables that were derived from deleted objects. Modifications are handled as deletions followed by creation of a “new” object, for which the functions are called.

You can do anything in those processing functions to decide what to do with this data. If you are not interested in that OSM object, simply return from the function. If you want to add the OSM object to some table call the add_row() (Version >= 1.7.0 or insert()) function on that table.

The parameter table (object) has the following fields and functions:

These are only available if the -x|--extra-attributes option is used and the OSM input file actually contains those fields.

The as_* functions will return a NULL geometry (check with is_null()) if the geometry can not be created for some reason, for instance a polygon can only be created from closed ways. This can also happen if your input data is incomplete, for instance when nodes referenced from a way are missing.

The as_linestring() and as_polygon() functions can only be used on ways.

Version >= 1.7.1 The as_multipoint() function can be used on nodes and relations. For nodes it will always return a point geometry, for relations a point or multipoint geometry with all available node members.

The as_multilinestring() and as_multipolygon() functions, on the other hand, can be used for ways and for relations. The latter will either return a linestring/polygon or a multilinestring/multipolygon, depending on whether the result is a single geometry or a multi-geometry.

If you need all geometries of a relation, you can use as_geometrycollection(). It will contain all geometries which can be generated from node and way members of the relation in order of those members. Members of type “relation” are ignored. Node members will result in a point geometry, way members will result in a linestring geometry. Geometries that can’t be built are missing. (You can detect this by counting the number of geometries with num_geometries() and comparing that to the number of members of type node and relation.) If no valid geometry can be created, so if the geometry collection would be empty, a null geometry is returned instead.

The add_row function

-- definition of the table:
table_pois = osm2pgsql.define_node_table('pois', {
    { column = 'tags', type = 'jsonb' },
    { column = 'name', type = 'text' },
    { column = 'geom', type = 'point' },
})
...
function osm2pgsql.process_node(object)
...
    table_pois:add_row({
        tags = object.tags,
        name = object.tags.name,
        geom = { create = 'point' }
    })
...
end

The add_row() function takes a single table parameter, that describes what to fill into all the database columns. Any column not mentioned will be set to NULL.

The geometry column is somewhat special. You have to define a geometry transformation that will be used to transform the OSM object data into a geometry that fits into the geometry column. See the next section for details. If you want more flexible geometry processing you need to use the insert() function available in version >= 1.70.

Note that you can’t set the object id, this will be handled for you behind the scenes.

Geometry Transformations for the add_row() Function

Currently these geometry transformations are supported:

• { create = 'point' }. Only valid for nodes, create a ‘point’ geometry.
• { create = 'line' }. For ways or relations. Create a ‘linestring’ or ‘multilinestring’ geometry.
• { create = 'area' } For ways or relations. Create a ‘polygon’ or ‘multipolygon’ geometry.

Some of these transformations can have parameters:

• The line transformation has an optional parameter split_at. If this is set to anything other than 0, long linestrings will be split up into parts no longer than this value.
• Version == 1.3.0 The area transformation has an optional parameter multi. If this is set to false (the default), a multipolygon geometry will be split up into several polygons. If this is set to true, the multipolygon geometry is kept as one.
• Version >= 1.4.0 The area transformation has an optional parameter split_at. If this is not set or set to nil (the default), (multi)polygon geometries will be kept as is, if this is set to 'multi' multipolygon geometries will be split into their polygon parts.

Note that in general it is your responsibility to make sure the column type of a geometry column can take the geometries created by the transformation. A point geometry column will not be able to store the result of an area transformation.

Version >= 1.4.0 If a transformation will result in a polygon geometry, but the column type of the geometry is multipolygon. The polygon will be “wrapped” automatically in a multipolygon to fit into the column. This can be useful when you want all your polygons and multipolygons to be of the same database type. Some programs handle columns of uniform geometry type better.

If no geometry transformation is set, osm2pgsql will, in some cases, assume a default transformation. These are the defaults:

• For node tables, a point column gets the node location.
• For way tables, a linestring column gets the complete way geometry, a polygon column gets the way geometry as area (if the way is closed and the area is valid).

The insert() Function

This function is only available in Version >= 1.7.0. Use add_row() in earlier versions.

-- definition of the table:
table_pois = osm2pgsql.define_node_table('pois', {
    { column = 'tags', type = 'jsonb' },
    { column = 'name', type = 'text' },
    { column = 'geom', type = 'point' },
})
...
function osm2pgsql.process_node(object)
...
    table_pois:insert({
        tags = object.tags,
        name = object.tags.name,
        geom = object:as_point()
    })
...
end

The insert() function takes a single table parameter, that describes what to fill into all the database columns. Any column not mentioned will be set to NULL. It returns true if the insert was successful.

Note that you can’t set the object id, this will be handled for you behind the scenes.

Handling of NULL in insert() Function

Any column not set, or set to nil (which is the same thing in Lua), or set to the null geometry, will be set to NULL in the database. If the column is defined with not_null = true in the table definition, the row will not be inserted. Usually that is just what you want, bad data is silently ignored that way.

If you want to check whether the insert actually happened, you can look at the return values of the insert() command. The insert() function actually returns up to four values:

local inserted, message, column, object = table:insert(...)

The last three are only set if the first is false.

Stages

When processing OSM data, osm2pgsql reads the input file(s) in order, nodes first, then ways, then relations. This means that when the ways are read and processed, osm2pgsql can’t know yet whether a way is in a relation (or in several). But for some use cases we need to know in which relations a way is and what the tags of these relations are or the roles of those member ways. The typical case are relations of type route (bus routes etc.) where we might want to render the name or ref from the route relation onto the way geometry.

The osm2pgsql flex output supports this use case by adding an additional “reprocessing” step. Osm2pgsql will call the Lua function osm2pgsql.select_relation_members() for each added, modified, or deleted relation. Your job is to figure out which way members in that relation might need the information from the relation to be rendered correctly and return those ids in a Lua table with the only field ‘ways’. This is usually done with a function like this:

function osm2pgsql.select_relation_members(relation)
    if relation.tags.type == 'route' then
        return { ways = osm2pgsql.way_member_ids(relation) }
    end
end

Instead of using the helper function osm2pgsql.way_member_ids() which returns the ids of all way members, you can write your own code, for instance if you want to check the roles.

Note that select_relation_members() is called for deleted relations and for the old version of a modified relation as well as for new relations and the new version of a modified relation. This is needed, for instance, to correctly mark member ways of deleted relations, because they need to be updated, too. The decision whether a way is to be marked or not can only be based on the tags of the relation and/or the roles of the members. If you take other information into account, updates might not work correctly.

In addition you have to store whatever information you need about the relation in your process_relation() function in a global variable.

After all relations are processed, osm2pgsql will reprocess all marked ways by calling the process_way() function for them again. This time around you have the information from the relation in the global variable and can use it.

If you don’t mark any ways, nothing will be done in this reprocessing stage.

(It is currently not possible to mark nodes or relations. This might or might not be added in future versions of osm2pgsql.)

You can look at osm2pgsql.stage to see in which stage you are.

You want to do all the processing you can in stage 1, because it is faster and there is less memory overhead. For most use cases, stage 1 is enough.

Processing in two stages can add quite a bit of overhead. Because this feature is new, there isn’t much operational experience with it. So be a bit careful when you are experimenting and watch memory and disk space consumption and any extra time you are using. Keep in mind that:

• All data stored in stage 1 for use in stage 2 in your Lua script will use main memory.
• Keeping track of way ids marked in stage 1 needs some memory.
• To do the extra processing in stage 2, time is needed to get objects out of the object store and reprocess them.
• Osm2pgsql will create an id index on all way tables to look up ways that need to be deleted and re-created in stage 2.

Type Conversions

The add_row() and insert() functions will try its best to convert Lua values into corresponding PostgreSQL values. But not all conversions make sense. Here are the detailed rules:

1. Lua values of type function, userdata, or thread will always result in an error.
2. The Lua type nil is always converted to NULL.
3. If the result of a conversion is NULL and the column is defined as NOT NULL, an error is thrown from the add_row() function. See above for NULL handling in insert().
4. The Lua type table is converted to the PostgreSQL type hstore if and only if all keys and values in the table are string values.
5. For boolean columns: The number 0 is converted to false, all other numbers are true. Strings are converted as follows: "yes", "true", "1" are true; "no", "false", "0" are false, all others are NULL.
6. For integer columns (int2, int4, int8): Boolean true is converted to 1, false to 0. Numbers that are not integers or outside the range of the type result in NULL. Strings are converted to integers if possible otherwise the result is NULL.
7. For real columns: Booleans result in an error, all numbers are used as is, strings are converted to a number, if that is not possible the result is NULL.
8. For direction columns (stored as int2 in the database): Boolean true is converted to 1, false to 0. The number 0 results in 0, all positive numbers in 1, all negative numbers in -1. Strings "yes" and "1" will result in 1, "no" and "0" in 0, "-1" in -1. All other strings will result in NULL.
9. Version >= 1.5.0 For json and jsonb columns string, number, and boolean values are converted to their JSON equivalent as you would expect. (The special floating point numbers NaN and Inf can not be represented in JSON and are converted to null). An empty table is converted to an (empty) JSON object, tables that only have consecutive integer keys starting from 1 are converted into JSON arrays. All other tables are converted into JSON objects. Mixed key types are not allowed. Osm2pgsql will detect loops in tables and return an error.
10. For text columns and any other not specially recognized column types, booleans result in an error and numbers are converted to strings.
11. Version >= 1.7.0 For insert() only: Geometry objects are converted to their PostGIS counterparts. Null geometries are converted to database NULL. Geometries in WGS84 will automatically be transformed into the target column SRS if needed. Non-multi geometries will automatically be transformed into multi-geometries if the target column has a multi-geometry type.

If you want any other conversions, you have to do them yourself in your Lua code. Osm2pgsql provides some helper functions for other conversions, see the Lua helper library (Appendix B).

Conversion to json and jsonb columns is only available from osm2pgsql 1.5.0 onwards. In versions before that you have to provide valid JSON from your Lua script to those columns yourself.

Geometry Objects in Lua

Version >= 1.7.0

Lua geometry objects are created by calls such as object:as_point() or object:as_polygon() inside processing functions. It is not possible to create geometry objects from scratch, you always need an OSM object.

You can write geometry objects directly into geometry columns in the database using the table insert() function. (You can not do that using the add_row() function, it has a different geometry handling, see above.) But you can also transform geometries in multiple ways.

Geometry objects have the following functions. They are modelled after the PostGIS functions with equivalent names.

The Lua length operator (#) returns the number of geometries in the geometry object, it is synonymous to calling num_geometries().

Converting a geometry to a string (tostring(geom)) returns the geometry type (as with the geometry_type() function).

All geometry functions that return geometries will return a NULL geometry on error. All geometry functions handle NULL geometry on input in some way. So you can always chain geometry functions and if there is any problem on the way, the result will be a NULL geometry. Here is an example:

local area = object:as_polygon():transform(3857):area()
-- area will be 0.0 if not a polygon or transformation failed

To iterate over the members of a multi-geometry use the geometries() function:

local geom = object:as_multipolygon()
for g in geom:geometries() do
    landuse.insert({
        geom = g,
        ...
    })
end

In Lua you can not get at the actual contents of the geometries, i.e. the coordinates and such. This is intentional. Writing functions in Lua that do something with the coordinates will be much slower than writing those functions in C++, so Lua scripts should concern themselves only with the high-level control flow, not the details of the geometry. If you think you need some function to access the internals of a geometry, start a discussion on Github.

The Pgsql Output

The pgsql output is the original output osm2pgsql started with. It was designed for rendering OpenStreetMap data, principally with Mapnik. It is still widely used although it is somewhat limited in how the data can be represented in the database.

If you are starting a new project with osm2pgsql, we recommend you use the flex output instead. The pgsql output will not receive all the new features that the flex output has or will get in the future. There are a lot of old configs and styles out there that need the pgsql output, though. So it will not go away anytime soon.

Database Layout

The pgsql output always creates a fixed list of four database tables shown below. The PREFIX can be set with the -p, --prefix option, the default is planet_osm. (Note that this option is also interpreted by the middle!)

All tables contain a geometry column named way (which is not the most intuitive name, but it can’t be changed now because of backwards compatibility).

Style File

Some aspects of how the pgsql output converts OSM data into PostgreSQL tables can be configured via a style file. The default style file that is usually installed with osm2pgsql is suitable for rendering many styles. It contains some documentation of the style syntax and works well as an example or starting point for your own changes. With a custom style file, you can control how different object types and tags map to different columns and data types in the database.

The style file is a plain text file containing four columns separated by spaces. As each OSM object is processed it is checked against conditions specified in the first two columns. If they match, processing options from the third and fourth columns are applied.

Possible values for the flags are:

Closed ways with tag area=yes and relations with type=multipolygon or type=boundary will be imported as polygons even if no polygon flag is set. Non-closed ways and closed ways with area=no will always be imported as lines.

Nodes are always placed in the “point” table, never in the “line” or “polygon” table.

Usually only objects which result in at least one of the columns declared in the style file being not NULL will be added to the database. But see below in the Use of Hstore section.

The style file may also contain comments. Any text between a # and the end of the line will be ignored.

Special ‘Tags’

There are several special values that can be used in the tag column (second column) of the style file for creating additional fields in the database which contain things other than tag values.

To use these, you must use the command line option --extra-attributes.

If importing with both --hstore and --extra-attributes the meta-data will end up in the tags hstore column regardless of the style file.

Schemas and Tablespaces

Usually all tables, indexes and functions that the pgsql output creates are in the default public schema and use the default tablespace.

Version >= 1.4.0 You can use the command line option --output-pgsql-schema=SCHEMA to tell osm2pgsql that it should use the specified schema to create the tables, indexes, and functions in. Note that you have to create the schema before running osm2pgsql and make sure the database user was granted the rights to create tables, indexes, and functions in this schema. Read more about schemas in the PostgreSQL documentation.

Sometimes you want to create special PostgreSQL tablespaces and put some tables or indexes into them. Having often used indexes on fast SSD drives, for instance, can speed up processing a lot. There are three command line options that the pgsql output interprets: To set the tablespace for output tables and indexes, use the --tablespace-main-data=TABLESPC and --tablespace-main-index=TABLESPC options, respectively. You can also use the -i, --tablespace-index=TABLESPC option which will set the index tablespace for the pgsql output as well as for the middle! Note that it is your job to create the tablespaces before calling osm2pgsql and making sure the disk behind it is large enough.

Coastline Processing

The natural=coastline tag is suppressed by default, even if you import the natural=* key. Many maps get the coastlines from a different source, so it does not need to import them from the input file. You can use the --keep-coastlines parameter to change this behavior if you want coastlines in your database. See the OSM wiki for more information on coastline processing.

Use of Hstore

Hstore is a PostgreSQL data type that allows storing arbitrary key-value pairs in a single column. It needs to be installed on the database with CREATE EXTENSION hstore;

Hstore is used to give more flexibility in using additional tags without reimporting the database, at the cost of less speed and more space.

By default, the pgsql output will not generate hstore columns. The following options are used to add hstore columns of one type or another:

• --hstore or -k adds any tags not already in a conventional column to a hstore column called tags. With the default stylesheet this would result in tags like highway appearing in a conventional column while tags not in the style like name:en or lanes:forward would appear only in the hstore column.

• --hstore-all or -j adds all tags to a hstore column called tags, even if they’re already stored in a conventional column. With the standard stylesheet this would result in tags like highway appearing in conventional column and the hstore column while tags not in the style like name:en or lanes:forward would appear only in the hstore column.

• --hstore-column or -z, which adds an additional column for tags starting with a specified string, e.g. --hstore-column 'name:' produces a hstore column that contains all name:xx tags. This option can be used multiple times.

You can not use both --hstore and --hstore-all together.

The following options can be used to modify the behaviour of the hstore columns:

• --hstore-match-only modifies the above options and prevents objects from being added if they only have tags in the hstore column and no tags in the non-hstore columns. If neither of the above options is specified, this option is ignored.

• --hstore-add-index adds indexes to all hstore columns. This option is ignored if there are no hstore columns. Using indexes can speed up arbitrary queries, but for most purposes partial indexes will be faster. You have to create those yourself.

Either --hstore or --hstore-all when combined with --hstore-match-only should give the same rows as no hstore, just with the additional hstore column.

Note that when you are using the --extra-attributes option, all your nodes, ways, and relations essentially get a few extra tags. Together with the hstore options above, the object attributes might end up in your hstore column(s) possibly using quite a lot of space. This is especially true for the majority of nodes that have no tags at all, which means they would normally not appear in your output tables. You might want to use --hstore-match-only in that case.

Lua Tag Transformations

The pgsql output supports Lua scripts to rewrite tags before they enter the database. Use the command line option --tag-transform-script=SCRIPT to enable this.

There is inevitably a performance hit with any extra processing. The Lua tag transformation is a little slower than the C++-based default, but this will probably not matter much in practice. But Lua pre-processing may save you further processing later. The Lua transformations allow you, for instance, to unify disparate tagging (for example, highway=path; foot=yes and highway=footway) and perform complex queries, potentially more efficiently than writing them as rules in your stylesheet which are executed at rendering time.

Note that this is a totally different mechanism than the Lua scripts used in the flex output.

The Lua script needs to implement the following functions:

function filter_tags_node(tags, num_tags)
return filter, tags

function filter_tags_way(tags, num_tags)
return filter, tags, polygon, roads

function filter_basic_tags_rel(tags, num_tags)
return filter, tags

These take a set of tags as a Lua key-value table, and an integer which is the number of tags supplied.

The first return value is filter, a flag which you should set to 1 if the way/node/relation should be filtered out and not added to the database, 0 otherwise. (They will still end up in the slim mode tables, but not in the rendering tables.)

The second return value is tags, a transformed (or unchanged) set of tags, which will be written to the database.

filter_tags_way returns two additional flags:

• poly should be 1 if the way should be treated as a polygon, or 0 if it is to be treated as a line.
• roads should be 1 if the way should be added to the roads table, 0 otherwise.

function filter_tags_relation_member(tags, member_tags, roles, num_members)
return filter, tags, member_superseded, boundary, polygon, roads

The function filter_tags_relation_member is more complex and can handle more advanced relation tagging, such as multipolygons that take their tags from the member ways.

This function is called with the tags from the relation; a set of tags for each of the member ways (member relations and nodes are ignored); the set of roles for each of the member ways; and the number of members. The tag and role sets are both arrays (indexed tables).

As with the other functions, it should return a filter flag, and a transformed set of tags to be applied to the relation in later processing. The third return value, member_superseded, is obsolete and will be ignored. The fourth and fifth return values, boundary and polygon, are flags that specify if the relation should be processed as a line, a polygon, or both (e.g. administrative boundaries). Set the final return value, roads, to 1 if the geometry should be added to the roads table.

There is a sample tag transform Lua script in the repository as an example, which (nearly) replicates the built-in transformations and can be used as a template for one’s own scripts.

Test your Lua script with small excerpts before applying it to a whole country or even the planet. Be aware that the Lua tagtransform allows to run arbitrary code on your system. Only run scripts from trusted sources!

Pgsql Output Command Line Options

These are all the command line options interpreted by the pgsql output:

Middle

This chapter is incomplete.

The middle keeps track of all OSM objects read by osm2pgsql and the relationships between those objects. It knows, for instance, which nodes are used by which ways, or which members a relation has. It also keeps track of all node locations. This information is necessary to build way geometries from way nodes and relation geometries from members and it is necessary when updating data, because OSM change files only contain changed objects themselves and not all the related objects needed for creating an object’s geometry.

Database Structure

The middle stores its data in the database in the following tables. The PREFIX can be set with the -p, --prefix option, the default is planet_osm. (Note that this option is also interpreted by the pgsql output!)

The names and structure of these tables, colloquially referred to as “slim tables”, are an internal implementation detail of osm2pgsql. While they do not usually change between releases of osm2pgsql, be advised that if you rely on the content or layout of these tables in your application, it is your responsibility to check whether your assumptions are still true in a newer version of osm2pgsql before updating. See this issue for a discussion of this topic.

Flat Node Store

--flat-nodes specifies that instead of a table in PostgreSQL, a binary file is used as a database of node locations. This should only be used on full planet imports or very large extracts (e.g. Europe) but in those situations offers significant space savings and speed increases, particularly on mechanical drives. The file takes approximately 8 bytes * maximum node ID, or more than 50 GiB, regardless of the size of the extract.

If you are using the --drop option, the flat node file will be deleted after import.

Caching

--cache specifies how much memory in MB to allocate for caching information. In --slim mode, this is just node positions while in non-slim mode it has to store information about ways and relations too. The rule of thumb in slim mode is as follows: use the size of the PBF file you are trying to import or about 75% of RAM, whatever is smaller. Make sure there is enough RAM left for PostgreSQL. It needs at least the amount of shared_buffers given in its configuration. You may also set --cache to 0 to disable node caching completely. This makes only sense when a flat node file is given and there is not enough RAM to fit most of the cache.

Bucket Index for Slim Mode

Version >= 1.4.0 This is only available from osm2pgsql version 1.4.0! It is enabled by default since 1.7.0.

Osm2pgsql can use an index for way node lookups in slim mode that needs a lot less disk space than earlier versions did. For a planet the savings can be about 200 GB! Because the index is so much smaller, imports are faster, too. Lookup times are slightly slower, but this shouldn’t be an issue for most people.

If you are not using slim mode and/or not doing updates of your database, this does not apply to you.

For backwards compatibility osm2pgsql will never update an existing database to the new index. It will keep using the old index. So you do not have to do anything when upgrading osm2pgsql.

If you want to use the new index, there are two ways of doing this: The “safe” way for most users and the “do-it-yourself” way for expert users. Note that once you switched to the new index, older versions of osm2pgsql will not work correctly any more.

Update for Most Users

Version >= 1.7.0 In versions 1.4.0 to 1.6.0 the index was not enabled by default. Add --middle-way-node-index-id-shift=5 as command line option for these versions. Do not use a number different than 5 unless you know what you are doing.

If your database was created with an older version of osm2pgsql you might want to start again from an empty database. Just do a reimport and osm2pgsql will use the new space-saving index.

Update for Expert Users

This is only for users who are very familiar with osm2pgsql and PostgreSQL operation. You can break your osm2pgsql database beyond repair if something goes wrong here and you might not even notice.

You can create the index yourself by following these steps:

Drop the existing index. Replace {prefix} by the prefix you are using. Usually this is planet_osm:

DROP INDEX {prefix}_ways_nodes_idx;

Create the index_bucket function needed for the index. Replace {way_node_index_id_shift} by the number of bits you want the id to be shifted. If you don’t have a reason to use something else, use 5:

CREATE FUNCTION {prefix}_index_bucket(int8[]) RETURNS int8[] AS $$
  SELECT ARRAY(SELECT DISTINCT unnest($1) >> {way_node_index_id_shift})
$$ LANGUAGE SQL IMMUTABLE;

Now you can create the new index. Again, replace {prefix} by the prefix you are using:

CREATE INDEX {prefix}_ways_nodes_bucket_idx ON {prefix}_ways
  USING GIN ({prefix}_index_bucket(nodes))
  WITH (fastupdate = off);

If you want to create the index in a specific tablespace you can do this:

CREATE INDEX {prefix}_ways_nodes_bucket_idx ON {prefix}_ways
  USING GIN ({prefix}_index_bucket(nodes))
  WITH (fastupdate = off) TABLESPACE {tablespace};

Id Shift (for Experts)

When an OSM node is changing, the way node index is used to look up all ways that use that particular node and therefore might have to be updated, too. This index is quite large, because most nodes are in at least one way.

When creating a new database (when used in create mode with slim option), osm2pgsql can create a “bucket index” using a configurable id shift for the nodes in the way node index. This bucket index will create index entries not for each node id, but for “buckets” of node ids. It does this by shifting the node ids a few bits to the right. As a result there are far fewer entries in the index, it becomes a lot smaller. This is especially true in our case, because ways often contain consecutive nodes, so if node id n is in a way, there is a good chance, that node id n+1 is also in the way.

On the other hand, looking up an id will result in false positives, so the database has to retrieve more ways than absolutely necessary, which leads to the considerable slowdown.

You can set the shift with the command line option --middle-way-node-index-id-shift. Values between about 3 and 6 might make sense, from some tests it looks like 5 is a good value.

To completely disable the bucket index and create an index compatible with earlier versions of osm2pgsql, use --middle-way-node-index-id-shift=0.

Middle Command Line Options

Expire

When osm2pgsql is processing OSM changes, it can create a list of (web mercator) tiles that are affected by those changes. This list can later be used to delete or re-render any changed tiles you might have cached. Osm2pgsql only creates this list. How to actually expire the tiles is outside the scope of osm2pgsql. Expire only makes sense in append mode.

Tile expiry will probably only work when creating web mercator (EPSG:3857) geometries. When using other output projections osm2pgsql can still generate expire files, but it is unclear how useful they are.

Tile expiry is only enabled if the -e, --expire-tiles option is set to anything but 0. Use this option to set the zoom level(s) which should be used for tile expiry. Details are below.

Details of the Expire Calculations

To figure out which tiles need to be expired, osm2pgsql looks at the old and new geometry of each changed feature and marks all tiles as expired that are touched by either or both of those geometries. For new features, there is only the new geometry, of course. For deleted features only the old geometry is used.

If the geometry of a feature didn’t change but the tags changed in some way that will always trigger the expiry mechanism. It might well be that the tag change does not result in any visible change to the map, but osm2pgsql can’t know that, so it always marks the tiles for expiry.

Which tiles are marked to be expired depends on the geometry type generated:

• For point geometries, the tile which contains this point is marked as expired.
• For line geometries (linestring or multilinestring) all tiles intersected by the line are marked as expired.
• For (multi)polygons all tiles in the bounding box are marked as expired. If the (multi)polygon is too large, only the boundary of the (multi)polygon is expired as described for line geometries above. “too large” in this case means with a bounding box of greater width or height (in meters) than what’s set with the --expire-bbox-size option (default is 20000).

In each case neighboring tiles might also be marked as expired if the feature is within a buffer of the tile boundary. This is so that larger icons, thick lines, labels etc. which are rendered just at a tile boundary which “overflow” into the next tile are handled correctly. Currently that buffer is set at 10% of the tile size, it is not possible to change it.

The Expire File

The list of tiles to be expired is written to a file. The name is specified with the -o, --expire-output=FILENAME option, it defaults to dirty_tiles in the current directory.

The generated expire file is a text file that contains one tile descriptor per line in the format ZOOM/X/Y. Tiles appear only once in the file even if multiple geometry changes are affecting the same tile.

If you set the -e, --expire-tiles option to a single zoom level, for instance --expire-tiles=12 then all tiles in the expire tile have that zoom level. If you set the option to a range of zoom levels, for instance -e 10-14, the file contains the tiles in all those zoom levels.

Memory requirements for osm2pgsql rise with the maximum zoom level used and the number of changes processed. This is usually no problem for zoom level 12 or 14 tiles, but might be an issue if you expire on zoom level 18 and have lots of changes to process.

Advanced Topics

Notes on Memory Usage

Importing an OSM file into the database is very demanding in terms of RAM usage. Osm2pgsql and PostgreSQL are running in parallel at this point and both need memory. You also need enough memory for general file system cache managed by the operating system, otherwise all IO will become too slow.

PostgreSQL blocks at least the part of RAM that has been configured with the shared_buffers parameter during PostgreSQL tuning and needs some memory on top of that. (See Tuning the PostgreSQL Server). Note that the PostgreSQL manual recommends setting shared_buffers to 25% of the memory in your system, but this is for a dedicated database server. When you are running osm2pgsql on the same host as the database, this is usually way too much.

Osm2pgsql needs at least 2GB of RAM for its internal data structures, potentially more when it has to process very large relations. In addition it needs to maintain a cache for node locations. The size of this cache can be configured with the parameter --cache.

When importing with a flatnode file (option --flat-nodes), it is best to disable the node cache completely (--cache=0) and leave the memory for the system cache to speed up accessing the flatnode file.

For imports without a flatnode file, set --cache approximately to the size of the OSM pbf file you are importing. (Note that the --cache setting is in MByte). Make sure you leave enough RAM for PostgreSQL and osm2pgsql as mentioned above. If the system starts swapping or you are getting out-of-memory errors, reduce the cache size or consider using a flatnode file.

When you are running out of memory you’ll sometimes get a bad_alloc error message. But more often osm2pgsql will simply crash without any useful message. This is, unfortunately, something we can not do much about. The operating system is not telling us that there is no memory available, it simply ends the program. This is due to something called “overcommit”: The operating system will allow the program to allocate more memory than there is actually available because it is unlikely that all programs will actually need this memory and at the same time. Unfortunately when programs do, there isn’t anything that can be done except crash the program. Memory intensive programs like osm2pgsql tend to run into these problems and it is difficult to predict what will happen with any given set of options and input files. It might run fine one day and crash on another when the input is only slightly different.

Note also that memory usage numbers reported by osm2pgsql itself or tools such as ps and top are often confusing and difficult to interpret. If it looks like you are running out of memory, try a smaller extract, or, if you can, use more memory, before reporting a problem.

Parallel Processing

Some parts of the osm2pgsql processing can run in parallel. Depending on the hardware resources of you machine, this can make things faster or slower or even overwhelm your system when too many things happen in parallel and your memory runs out. For normal operation the defaults should work, but you can fine-tune the behaviour using some command line options.

Osm2pgsql will do some of its processing in parallel. Usually it will use however many threads your CPUs support, but no more than 4. For most use cases this should work well, but you can tune the number of threads used with the --number-processes command line option. (Note that the option is a bit of a misnomer, because this sets the number of threads used, they are all in a single process.) If disks are fast enough e.g. if you have an SSD, then this can greatly increase speed of the “going over pending ways” and “going over pending relations” stages on a multi-core server. Past 8 threads or so this will probably not gain you any speed advantage.

By default osm2pgsql starts the clustering and index building on all tables in parallel to increase performance. This can be a disadvantage on slow disks, or if you don’t have enough RAM for PostgreSQL to do the parallel index building. PostgreSQL potentially needs the amount of memory set in the maintenance_work_mem config setting for each index it builds. With 7 or more indexes being built in parallel this can mean quite a lot of memory is needed. Use the --disable-parallel-indexing option to build the indexes one after the other.

Handling of Forward Dependencies

Whenever a node changes, osm2pgsql will find all ways and relations that have this node as a member and reprocess them. Similarly, whenever a way changes, their parent relations are reprocessed. So osm2pgsql will automatically handle “forward dependencies” from nodes and ways to their parents. Almost always this is the behaviour you want.

Version >= 1.4.0 This behaviour can be disabled with the command line option --with-forward-dependencies=false. It is used by Nominatim which uses the specialized Gazetteer output which doesn’t need this behaviour.

Using Database While Updating

To improve performance osm2pgsql uses several parallel threads to import or update the OSM data. This means that there is no transaction around all the updates. If you are querying the database while osm2pgsql is running, you might be able to see some updates but not others. While an import is running you should not query the data. For updates it depends a bit on your use case.

In most cases this is not a huge problem, because OSM objects are mostly independent of one another. If you are

• writing OSM objects into multiple tables,
• using two-stage processing in the Flex output, or
• doing complex queries joining several tables

you might see some inconsistent data, although this is still rather unlikely. If you are concerned by this, you should stop any other use of the database while osm2pgsql is running. This is something that needs to be done outside osm2pgsql, because osm2pgsql doesn’t know what else is running and whether and how it might interact with osm2pgsql.

Note that even if you are seeing inconsistent data at some point, the moment osm2pgsql is finished, the data will be consistent again. If you are running a tile server and using the expire functionality, you will, at that point, re-render all tiles that might be affected making your tiles consistent again.

Clustering by Geometry

Typical use cases for a database created by osm2pgsql require querying the database by geometry. To speed up this kind of query osm2pgsql will by default cluster the data in the output tables by geometry which means that features which are in reality near to each other will also be near to each other on the disk and access will be faster.

Version >=1.7.0 If a table has multiple geometry columns, clustering will always be by the first geometry column.

This clustering is achieved by ordering and copying each whole table after the import. This will take some time and it means you will temporarily need twice the disk space.

When you are using the flex output, you can disable clustering by setting the cluster table option to no (see the Advanced Table Definition section).

Tips & Tricks

The flex output allows a wide range of configuration options. Here are some extra tips & tricks.

Using create_only Columns for Postprocessed Data

Sometimes it is useful to have data in table rows that osm2pgsql can’t create. For instance you might want to store the center of polygons for faster rendering of a label.

To do this define your table as usual and add an additional column, marking it create_only. In our example case the type of the column should be the PostgreSQL type GEOMETRY(Point, 3857), because we don’t want osm2pgsql to do anything special here, just create the column with this type as is.

polygons_table = osm2pgsql.define_area_table('polygons', {
    { column = 'tags', type = 'hstore' },
    { column = 'geom', type = 'geometry' },
    { column = 'center', sql_type = 'GEOMETRY(Point, 3857)', create_only = true },
    { column = 'area', type = 'area' },
})

After running osm2pgsql as usual, run the following SQL command:

UPDATE polygons SET center = ST_Centroid(geom) WHERE center IS NULL;

If you are updating the data using osm2pgsql --append, you have to do this after each update. When osm2pgsql inserts new rows they will always have a NULL value in center, the WHERE condition makes sure that we only do this (possibly expensive) calculation once.

Creating GENERATED Columns

From version 12 on PostgreSQL supports GENERATED columns. You can use these from osm2pgsql with a trick, by adding some “magic” wording to the sql_type in a column definition.

We don’t promise that this trick will work forever. Putting anything into the sql_type but a PostgreSQL type is strictly experimental.

It is probably best explained with an example. With this config you create a polygon table with an additional column center that is automatically filled with the centroid of the polygon added:

local polygons = osm2pgsql.define_area_table('polygons', {
    { column = 'tags', type = 'jsonb' },
    { column = 'geom', type = 'polygon', not_null = true },
    { column = 'center', create_only = true, sql_type = 'geometry(point, 3857) GENERATED ALWAYS AS (ST_Centroid(geom)) STORED' },
})

function osm2pgsql.process_way(object)
    if object.is_closed then
        polygons:insert({
            tags = object.tags,
            geom = object:as_polygon()
        })
    end
end

You could do this specific case simpler and faster by using the centroid() function in Lua. But PostGIS has a lot more interesting functions that you can use this way. Be sure to read the PostgreSQL documentation about the syntax and functionality of the GENERATED columns.

Make sure you have the create_only = true option set on the column, otherwise this will not work.

Accessing Environment Variables from Lua

In Lua scripts you can access environment variables with os.getenv("VAR"). The Lua config scripts that osm2pgsql uses are normal Lua scripts, so you can do that there, too.

Logging Memory Use of Lua Scripts

Lua scripts can use quite a lot of memory if you are not careful. This is usually only a problem when using two-stage processing. To monitor how much memory is currently used, use this function:

function create_memory_reporter(filename, frequency)
    local counter = 0
    local file = io.open(filename, 'w')
    file:write('timestamp,counter,mbyte\n')

    return function()
        if counter % frequency == 0 then
            local mem = collectgarbage('count')
            local ts = os.date('%Y-%m-%dT%H:%M:%S,')
            file:write(ts .. counter .. ',' .. math.ceil(mem / 1024) .. '\n')
            file:flush()
        end
        counter = counter + 1
    end
end

Then use it to create one or more memory_reporters. The first argument is the output file name, the second specifies after how many process callbacks it should trigger the output. Make sure this number is not too small, otherwise processing will become quite slow.

Here is an example use for the process_node callback:

local mr = create_memory_reporter('/tmp/osm2pgsql-lua-memlog.csv', 10000)

function osm2pgsql.process_node(object)
    mr()
    ...
end

You can have one memory reporter for nodes, ways, and relations together or have separate ones.

Checking Lua Scripts

If you have the Lua compiler (luac) installed (it comes with the installation of the Lua interpreter), you can use it to syntax check your Lua scripts independently of osm2pgsql. Just run luac -p SCRIPT.lua.

A script that fails this check will not work with osm2pgsql. But it is only a syntax check, it doesn’t run your script, so the script can still fail when used in osm2pgsql. But it helps getting rid of the simple errors quickly.