14 KiB

Raw Blame History

Installing Pelias

Mapzen offers the Mapzen Search service in hopes that as many people as possible will use it, but we also encourage people to set up their own Pelias instance.

For most cases, it's useful to have much of the installation process automated, so we suggest looking at the Pelias Vagrant image.

However, for more in-depth usage, to learn more about the working of Pelias, or to contribute back, manual setup is useful. These instructions will help you install Pelias from scratch manually.

Installation Overview

The steps for fully installing Pelias look like this:

Decide which datasets and settings will be used
Download appropriate data
Download Pelias code, using the appropriate branches
Set up Elasticsearch
Install the Elasticsearch schema using pelias-schema
Use one or more importers to load data into Elasticsearch
Start the API server to begin handling queries

System Requirements

In general, Pelias will require:

A working Elasticsearch 2.3 cluster. It can be on a single machine or across several
Node.js 0.12 or newer (Node 4 or 6 is recommended)
Up to 100GB disk space to download and extract data
Lots of RAM, 8GB is a good minimum. A full North America OSM import just fits in 16GB RAM

Choose your datasets

Pelias can currently import data from four different sources. The contents and description of these sources are available on our data sources page. Here we'll just focus on what to download for each one.

Who's on First

The Who's on First importer contains code and instructions for downloading WOF data.

Alternatively, there are two other ways to download Who's on First data. The first is to use the pre-created bundles. These consist of a series of archives that can be easily extracted (instructions are on the page).

For more advanced uses, or to contribute back to Who's on First, use the whosonfirst-data Github repository. Again, there are instructions. Note that this repo requires git-lfs, a lot of bandwidth, and 27GB (currently) of disk space.

Geonames

The pelias/geonames importer contains code and instructions for downloading Geonames data automatically. Individual countries, or the entire planet (1.3GB compressed) can be specified.

OpenAddresses

The OpenAddresses project includes numerous download options, all of which are .zip downloads. The full dataset is just over 6 gigabytes compressed, but there are numerous subdivision options. In any case, the .zip files simply need to be extracted to a directory of your choice, and Pelias can be configured to either import every .csv in that directory, or only selected files.

OpenStreetMap

OpenStreetMap has a nearly limitless array of download options, and any of them should work as long as they're in PBF format. Generally the files will have the extension .osm.pbf. Good sources include the Mapzen Metro Extracts (which has popular cities available immediately, or custom areas that take only a few minutes to build), and planet files listed on the OSM wiki. A full planet PBF file is about 36GB.

Choose your import settings

There are several options that should be discussed before starting any data imports, as they require a compromise between import speed and resulting data quality and richness.

Admin Lookup

Most data that is imported by Pelias comes to us incomplete: many data sources don't supply what we call admin hierarchy information: the neighbourhood, city, country, or other region that contains the record. In OpenAddresses, for example, many records contain only a housenumber, street name, and coordinates.

Fortunately, Who's on First contains a well-developed set of geometries for all admin regions from the neighbourhood to continent level. Through point-in-polygon lookup, our importers can derive this information!

The downsides to enabling admin lookup are increased memory requirements and longer import times. Because geometry data is quite large, expect to use about 6GB of RAM (not disk) during import just for this geometry data. And because of the complexity of the required calculations, imports with admin lookup are up to 10 times slower than without.

Who's on First, of course, always includes full hierarchy information because it's built into the dataset itself, so there's no tradeoff to be made. Who's on First data will always import quite fast and with full hierarchy information.

Address Deduplication

OpenAddresses data contains lots of addresses, but it also contains lots of duplicate data. To help reduce this problem we've built an address-deduplicator that can be run at import. It uses the OpenVenues deduplicator to remove records that are near each other and have names that are likely to be duplicates. Note that it's considerably smarter than simply doing exact comparisons of names and coordinates: it uses Geohash prefixes to compare nearby records, and the libpostal address normalizer to compare names, so it can tell that records with 101 Main St and 101 Main Street are likely to refer to the same place.

Unfortunately, our current implementation is very slow, and requires about 50GB of scratch disk space during a full planet import. It's worth noting that Mapzen Search currently does not deduplicate any data, although we hope to improve the performance of deduplication and resume using it eventually.

Considerations for full-planet builds

As may be evident from the dataset section above, importing all the data in all four supported datasets is worthy of its own discussion. Current full planet builds weigh in at over 320 million documents, and require about 230GB total storage in Elasticsearch. Needless to say, a full planet build is not likely to succeed on most personal computers.

Fortunately, because of services like AWS and the scalability of Elasticsearch, full planet builds are possible without too much extra effort. To set expectations, a cluster of 4 r3.xlarge AWS instances running Elasticsearch, and one c4.8xlarge instance running the importers can complete a full planet build in about two days.

Choose your Pelias code branch

As part of the setup instructions below, you'll be downloading several Pelias packages from source on Github. All of these packages offer 3 branches for various use cases. Based on your needs, you should pick one of these branches and use the same one across all of the Pelias packages.

production: contains only code that has been tested against a full-planet build and is live on Mapzen Search. This is the "safest" branch and it will change the least frequently, although we generally release new code at least once a week.

staging: these branches contain the code that is currently being tested against a full planet build for imminent release to Mapzen Search. It's useful to track what code will be going out in the next release, but not much else.

master: master branches contain the latest code that has passed code review, unit/integration tests, and is ready to be included in the next release. While we try to avoid it, the nature of the master branch is that it will sometimes be broken. That said, these are the branches to use for development of new features.

Installation

Download the Pelias repositories

At a minimum, you'll need the Pelias schema and api repositories, as well as at least one of the importers. Here's a bash snippet that will download all the repositories (they are all small enough that you don't have to worry about the space of the code itself), check out the production branch (which is probably the one you want), and install all the node module dependencies.

for repository in schema api whosonfirst geonames openaddresses openstreetmap; do
	git clone git@github.com:pelias/${repository}.git
	pushd $repository > /dev/null
	git checkout production # or staging, or remove this line to stay with master
	npm install
	popd > /dev/null
done

Customize Pelias Config

Nearly all configuration for Pelias is driven through a single config file: pelias.json. By default, Pelias will look for this file in your home directory, but you can configure where it looks. For more details, see the pelias-config repository.

The two main things of note to configure are where on the network to find Elasticsearch, and where to find the downloaded data files.

Pelias will by default look for Elasticsearch on localhost at port 9200 (the standard Elasticsearch port).

By taking a look at the default config, you can see the Elasticsearch configuration looks something like this:

{
  "esclient": {
  "hosts": [{
    "host": "localhost",
    "port": 9200
  }]

  ... // rest of config
}

If you want to connect to Elasticsearch somewhere else, change localhost as needed. You can specify multiple hosts if you have a large cluster. In fact, the entire esclient section of the config is sent along to the elasticsearch-js module, so any of its configuration options are valid.

The other major section, imports, defines settings for each importer. The defaults look like this:

{
 "imports": {
    "geonames": {
      "datapath": "./data",
      "adminLookup": false
    },
    "openstreetmap": {
      "datapath": "/mnt/pelias/openstreetmap",
	  "adminLookup": false,
      "leveldbpath": "/tmp",
      "import": [{
        "filename": "planet.osm.pbf"
      }]
    },
    "openaddresses": {
      "datapath": "/mnt/pelias/openaddresses",
      "adminLookup": false,
      "files": []
    },
    "whosonfirst": {
      "datapath": "/mnt/pelias/whosonfirst"
    }
  }
}

As you can see, the default datapaths are meant to be changed. This is also where you can enable admin lookup by overriding the default value.

Install Elasticsearch

Other than requiring Elasticsearch 2.3, nothing special in the Elasticsearch setup is required for Pelias, so please refer to the official 2.3 install docs.

Older versions of Elasticsearch are not supported.

Make sure Elasticsearch is running and connectable, and then you can continue with the Pelias specific setup and importing. Using a plugin like head or Marvel can help monitor Elasticsearch as you import data.

If you're using a terminal, you can also search and/or monitor Elasticsearch using their APIs.

Note: On large imports, Elasticsearch can be very sensitive to memory issues. Be sure to modify it's heap size from the default confiration to something more appropriate to your machine.

Set up the Elasticsearch Schema

The Elasticsearch Schema is analogous to the layout of a table in a traditional relational database, like MySQL or PostgreSQL. While Elasticsearch attempts to auto-detect a schema that works when inserting new data, this generally leads to non-optimal results. In the case of Pelias, inserting data without first applying the Pelias schema will cause all queries to fail completely: Pelias requires specific configuration settings for both performance and accuracy reasons.

Fortunately, now that your pelias.json file is configured with how to connect to Elasticsearch, the Schema repository can automatically create the Pelias index and configure it exactly as needed:

cd schema # assuming you've just run the bash snippet to download the repos from earlier
node scripts/create_index.js

If you want to reset the schema later (to start over with a new import or because the schema code has been updated), you can drop the index and start over like so:

# !! WARNING: this will remove all your data from pelias!!
node scripts/drop_index.js # it will ask for confirmation first
node scripts/create_index.js

Note that Elasticsearch has no analogy to a database migration, so you generally have to delete and reindex all your data after making schema changes.

Run the importers

Now that the schema is set up, you're ready to begin importing data!

Our goal is that eventually you'll be able to run all the importers with simply cd $importer_directory; npm start. We are now really close, and all but one importer follows this pattern!

That importer is the Geonames importer, please see its README file for the most up to date instructions. By the way, we'd love to see a pull request to allow it to read configuration from pelias.json like the other impoters.

Depending on how much data you've imported, now may be a good time to grab a coffee. Without admin lookup, the fastest speeds you'll see are around 10,000 records per second. With admin lookup, expect around 800-2000 inserts per second.

Start the API

As soon as you have any data in Elasticsearch, you can start running queries against the Pelias API server.

Again thanks to pelias.json, the API already knows how to connect to Elasticsearch, so all that's required to star the API is npm start. You can now send queries to http://localhost:3100/!

14 KiB Raw Blame History