Pelias is a set of tools for importing OpenStreetMap data into Elasticsearch, and a simple server to handle queries and autocomplete suggestions.

• PostgreSQL: You'll need a postGIS-enabled database with OpenStreetMap data, imported with osm2pgsql. NOTE: The import process expects certain fields, so you'll need to use the style file here: config/osm2pgsql.style
• Elasticsearch: For Search download the latest version of Elasticsearch
• Redis: Geonames lookup for quattroshapes cross-referencing
• Sidekiq: Used for background processing (also uses Redis)
• Ruby >= 2

To get set up, run the following.

$ git clone [email protected]:mapzen/pelias.git && cd pelias
$ bundle

$ bundle exec rake synonyms:build

$ bundle exec rake index:create

Geonames provide nice alternative names and populations for locations. We cross-reference this data with quattroshapes in the next step in order to provide a better search experience.

$ bundle exec rake geonames:prepare

These are shapes for various administrative shapes. They are provided by the http://quattroshapes.com/ project.

NOTE: These tasks are enqueued via Sidekiq and must be run in isolated steps. You can run them inline by using the environment variable ES_INLINE=1.

$ bundle exec rake quattroshapes:prepare_all
$ bundle exec rake quattroshapes:populate_admin0 ES_INLINE=1
$ bundle exec rake quattroshapes:populate_admin1 ES_INLINE=1
$ bundle exec rake quattroshapes:populate_admin2 ES_INLINE=1
$ bundle exec rake quattroshapes:populate_local_admin ES_INLINE=1
$ bundle exec rake quattroshapes:populate_locality ES_INLINE=1
$ bundle exec rake quattroshapes:populate_neighborhood ES_INLINE=1

Assuming you've set up a postGIS-enabled database with OSM data, the following will add all streets and addresses to the index, reverse geocoding them into the above shapes.

$ bundle exec rake osm:populate_street
$ bundle exec rake osm:populate_address
$ bundle exec rake osm:populate_poi

$ unicorn

You should now be able to access the server at http://localhost:8080/suggest?query=party

The following is a brief synopsis of setting up this environment including: approximate times to complete each step, amount of data, number of documents, etc.

• PostgreSQL/PostGIS: 1 c3.8xlarge
  • this is only to facilitate the fastest of initial load times into pelias
• Elasticsearch: 20 m3.2xlarge
  • optimization work to be done to lessen on heap storage requirements
  • assumes 80 shards, 1 replica per shard, half of physical memory allocated to ES for heap
• Sidekiq: 8 c1.medium
  • only required for initial import to complete in a timely manner
  • can be removed once complete or scaled back as required for updates on an ongoing basis

Using this hardware allocation, we also recommend the following during the initial data load:

• disable replication in elasticsearch
• set the index refresh interval to something north of an hour (or disable it altogether for the duration of the indexing process)
• in PostgreSQL, add the following index (this will take some time if you're working with a full planet installation):
  • CREATE INDEX limit_street_line ON planet_osm_line (name, highway);

Using the above architecture, we've observed the following load times:

• geonames + quattroshapes: roughly an hour
• osm: ~3 days

Documents in Elasticsearch upon completion of load:

• ~66 million

Unique data size on disk:

• ~300GB
  • ~600GB with one replica

This is our search endpoint. This is used to search the index for addresses, POIs, etc.

/search?query=brooklyn
/search?query=brooklyn&center=-74.08,40.77
/search?query=brooklyn&viewbox=-74.08,40.77,-73.9,40.67
/search?query=brooklyn&viewbox=-74.08,40.77,-73.9,40.67

This is an autocomplete suggestion endpoint. It provides search suggestions given text to look up.

/suggest?query=bro
/suggest?query=bro&size=5

This is the reverse geocoding endpoint. It takes lng and lat params and returns GeoJSON corresponding to the given location.

/reverse?lng=1&lat=2

Check out our demo here: http://mapzen.com/pelias

MIT License. See included LICENSE