These are the release notes for Pelias, the open source project. It includes news on the latest code changes as well as new features. There are also release notes specifically for the hosted instance of Pelias run by Mapzen, Mapzen Search.
- The
/placeendpoint would sometimes return records from a different layer than the one you asked for, but no longer! - Asking the
/reverseendpoint to return venues from Geonames used to return an error message. However, it should work, and now it does!
- Our Placeholder and interpolation services now have access logs. This should make it much easier to diagnose any issues and measure traffic
- The Point in Polygon service now can download data using the Who's on First downloader directly. It's a bit easier to set up the PIP service as a result!
- You can now specify the network interface the Pelias API will listen on, by setting the
HOSTenvironment variable. The default is still Node.js's default to listen on all interfaces. - It's now easy to tell which parser was used for a query, as the API will set
parser: "libpostal"when libpostal was used
This release marks a major update of our Placeholder service that is in charge of managing administrative area names. All sorts of queries that need to understand the name of a country, city, county, neighbourhood, or any other administrative area are improved! Especially improved should be queries that use non-English languages. The full details of the code changes are in the Placeholder pull request.
Over the last few days we have fixed several regressions that could cause bad responses from, and sometimes even crash, our API server.
- Thanks to @rabidllama's excellent PR, we now support queries with postalcode and region/country combinations.
- We've beefed up downloaders for OpenAddresses, OpenStreetMap, and Who's On First to download subsets of data.
- The alpha3 flag for country codes have been obsolete for quite some time now but we've removed all vestiges of it.
- Some OpenAddresses records have lon/lat at 0/0 meaning that, unless you have a ocean-going yacht and pretty sturdy jet ski, you can't get to addresses there (note to self: start "Pelias for Yachts"). We've now excluded those from the build.
- We now support the ability to specify the admin hierarchy part of an address query in any language (as long as a name in that language can be found in WOF), for example searching for 30 w 26th st, Нью Йорк, 미국 works like magic!
- We've added a warning when unexpected query parameters are encountered! This is a big deal because it helps catch misspellings of query parameter names, such as
layervslayers... we've all been there at least once.
- We found an invalid warning when falling back to a coarse-reverse geocoding strategy about
boundary.circle.radiusnot being supported, even though the query didn't specify a value for that parameter. We removed the unwanted warning. If you see that warning going forward, consider it valid and revise your query. - We fixed our production configuration to fully enable language headers (they were only partially available previously). Ooops!
- To support the ability to search for addresses using many languages, we've refactored a lot of our
apiexecution flow for address queries as well as the query logic sent to Elasticsearch- the new flow goes like this:
libpostal->placeholder->query-with-wof-ids-instead-of-placenames
- the new flow goes like this:
- We finally deleted some old unused code in the
api, such as the olddetails sanitizer - Logging for the microservice-wrapper module used in the
apihas been improved - We have added the ability to specify a default focus point within the
pelias-configfor theapito allow custom instances of Pelias to always bias results to a specified location - After 2 years since the release of
v1of our API, we've finally done away with the legacy URL configuration and no longer support/or any endpoint not containing the/v1prefix - All the
apisanitizers have been updated for consistency - We got rid of the
alpha3property in our schema and model... we're not even sure why we needed it anymore, so good riddance!
Today's release has no new code, but it's a big milestone nonetheless. For the first time, we are indexing over 500 million (that's half a billion) records, all from open data! We're grateful to all the work from contributors of OpenStreetMap, OpenAddresses, Who's on First, and Geonames that have made this possible and are looking forward to reaching the big one billion mark soon!
Our new Placeholder service queries were working great, except they weren't returning confidence scores. Now they have confidence scores just like all our other queries.
- After several months of work we have finally released massively improved admin queries! These use our new service, Placeholder. Placeholder knows all about cities, countries, and every other type of administrative area, and knows the relationships between them, so it can be used to return much better results.
- Our importers now have a higher Node.js memory limit. This should fix issues that were being caused by the ever-increasing amount of open-data available to our admin lookup service.
- We discovered that we were importing every jetway from OpenStreetMap, which isn't really something anyone wants to geocode, so we've improved our whitelists for what's allowed. Results from OSM should be a little cleaner now.
- Autocomplete now works correctly on very long names.
- Deduper will now prefer results with postalcodes, which comes to us courtesy of our friend @kevincennis.
- WOF Macrohoods have been added to the list of supported hierarchies.
- Speaking of updated dependencies, we switched to the latest release of addressit, which has a fix from, AGAIN, our friend @kevincennis! Hey, @kevincennis, you are awesome and we appreciate you! Thanks!
- We have removed the
alpha3field from our schema because it was basically the appendix of Pelias, in that it was leftover from a long time ago and is no longer needed for proper functioning of the system. This data can still be found in each record, but only in theparent.country_afield. - We've added dependency checks to precommit hooks and travis builds for all repositories.
- Admin-lookup was refactored and all importers have been updated to support the new interface. As a result, a lot of dependencies have been updated.
- WOF importer module has been updated to be more generic so it could be used in other importers and packages where we need to download and/or manipulate WOF data.
- We've also added more stringent checks for WOF records visiting Null Island, so they don't make it into our index.
- We've updated the fields we use for
populationdata from WOF after consulting with the data team. - The minimum size of results parameter we send to Elasticsearch has been locked down to 20, which ensures that we get back enough results to cut away results that aren't granular enough and dedupe.
- Our first big ticket item is technically a new feature, a code level change, and a bug fix all in one! We've created a standalone microservice whose job it is to handle point-in-polygon requests. So with this release, all reverse queries specifying admin layers will be directed to this new service, instead of going to Elasticsearch like it used to. As a user, you won't see any difference in the interface to these types of requests and you don't have to take any action to use the new functionality. However, faster and better results will be apparent!
- Our second big ticket item (we know, 2 in one release is awesome sauce!) is the long awaited upgrade to
libpostal 1.0. This is again a code level change that doesn't have any user interface implications but yield significant improvements in results. We can tell just by the number of old issues we were able to resolve as a result of this upgrade that this is a big moment for the Pelias engine. High-fives all around! - You know how we started supporting search queries with only postalcodes in them, like
/v1/search?text=90210? Well get excited, because we've added the ability to handle postalcode only queries instructuredsearch as well! So queries like/v1/search/structures?postalcode=90210will now work. More info here.
- For those hosting their own instances of Pelias, the new point-in-polygon service will have some impact on your setup. You can see the pip-service README.md for setup details and don't forget to add the url to your new pip service instance to the API section of your Pelias config file. If you don't specify the
pipServiceproperty in config, things will continue to work as they have before and Elasticsearch queries will continue to be used for coarse reverse requests.
{
"api": {
"host": "localhost:3100/v1/",
"textAnalyzer": "libpostal",
"pipService": "url to your pip service instance"
},
...
}
- All the code is in place for checking the language headers in requests and returning results with admin names in the specified language. The functionality is not yet enabled and requires a link to a
placeholderservice to work as expected. - We've swapped the order of operation in the case of interpolation queries to deduplicate after interpolation has taken place, not before. When deduplicating before interpolation, street segments were getting removed as duplicates before they could be sent to the interpolation service for evaluation, causing us to miss some addresses.
- We fixed a few minor bugs related to address interpolation. There were cases where the results had a mix of street centroids and addresses and the correct address was not showing up first. More details here.
- There was an issue with geonames admin records having incorrect ids in their admin hierarchy properties. They were basically masquerading as Who's on First ids leading to invalid results and general chaos. Well no more. We fixed it.
- We've added postalcodes to the Who's on First import process and enabled the postalcode-only query type, so users can now find postalcodes directly!
- This update required additions to the Pelias schema as well as query building logic.
- The download script in pelias-whosonfirst importer has also been updated to pull down postalcode data by default, so be aware when using that download command and when running the importer.
- This also means Who's on First imports will now take roughly 3 times as long, if importing the whole world... Yeah, there are a LOT of postalcodes in the world!
Pelias <3 Postalcodes
- Refactored pelias-wof-admin-lookup and updated to this latest version in all importers.
- This is in preparation for the upcoming fixes to our coarse reverse lookups.
- Next step will be to direct the API to this new Point-in-Polygon service when available to get more accurate reverse geocoding results.
- Geonames importer interface has finally been updated to match our other importers.
- Dependencies should now have the proper alpha3 ISO codes of their own in the country abbreviation (
country_a) properties, instead of alpha2 of the parent country as it did previously. See San Juan, PR for example. - Washington DC wasn't getting a region abbreviation at all, but that's water under the Arlington Memorial Bridge now!
- We're working on some fancy new things in dev so our only new feature released this week is retrying timed-out elasticsearch requests. By default, Pelias will retry a timed-out request up to 3 times before failing for good, though this number can be overridden by specifying
api.requestRetriesin your Pelias configuration.
- Big news! 🐯 we have soft launched our new street interpolation service which includes TIGER data. This allows us to return more address results than before. For more info see https://github.com/pelias/interpolation
- We've fixed a bug where structured queries would always return 'fallback' as the 'match_type'.
- The
/v1/structuredendpoint now supports thevenueparameter, which allows for searching for venues with a particular name. - We've improved result balance when using
focus.pointin the autocomplete endpoint. In particular, searching for cities far away from the focus point should work much better. More improvements tofocus.pointare planned for the near future.
- We've finished adding Pelias config validation (for example here in the Who's on First importer) to all of our repositories. It should be much harder for simple config errors such as typos to go undetected!
- The checks run before dropping data using pelias-schema have been improved. Accidentally deleting all your production data should now be even more difficult than before!
- We are now considerably more strict about read errors when loading Who's on First data. Any read errors, whether due to missing files or invalid JSON will cause the entire importer to fail immediately. This prevents builds from reporting succesful completion when some data is actually missing.
- The Elasticsearch
index_concurrencysetting has been removed. In theory this setting allowed tuning imports for better performance, but in practice it had little effect. It's going away in Elasticsearch 5, so this will make the transition that much easier.
- Searches for
St Louis, MOandSaint Louis, MOnow return the same thing (the same goes for towns starting withMount/MtandFort/Ft)
- Structured geocoding no longer fails horribly when the
addressparameter consists of only a house number
- We've released what was previously referred to as component geocoding in the new structured geocoding endpoint! It lives at
/v1/search/structured - We fixed a bug where specifying the same parameter twice (eg
/v1/search?text=paris&sources=geonames&sources=gn) would cause a 500 error. It now returns a helpful 400 error message that includes which parameter is duplicated, so that the request can be fixed. - Other errors that should have been 500 errors were being returned with status code 400. Fixing this will allow us to more quickly catch any 500 errors that happen in the future.
- We've just released beta support for component geocoding so instead of passing in a single input to the
/v1/searchendpoint, the parts of an address can be sent to/v1/beta/component! An example of this isaddress=201+Spear+St&locality=San+Francisco®ion=CA. We haven't officially named this geocoding type yet, so if you have a naming suggestion, please weigh in here! Our basic design doc for using this new beta feature is here, please check it out. We're still working out the final implementation (why it's currently deployed to our/v1/betatest bed) so check it out and don't hesitate to raise any issues you might encounter. Check out the acceptance tests for some more examples.
- We're enabling support for more response scenarios from libpostal! This release we're adding support for city+country, so requests for Paris, France and Reykjavík, Iceland are a lot cleaner.
- Speaking of Reykjavík, Iceland, support for inputs containing diacritics has improved. Now whether the input is Reykjavík, Iceland or Reykjavik, Iceland, results should be the same.
- Whether your input contains a 2- or 3-character ISO country code (
FRAvsFR), we'll find it!
- The
/v1/autocompleteendpoint now supports boundary.rect just like/v1/search - Labels for administrative areas should be improved in a few cases
- All Pelias code now requires Node.js 4.0 or newer! This means we can start using ES6 features.
letus have a party to celebrate! 🎉 - Label generation code has been extracted into pelias-labels
- Elasticsearch settings have been moved out of a dark corner and into pelias/config
- The whosonfirst importer has a better method for calculating hierarchy information. It's simpler and should be more accurate and faster!
- libpostal, the super-sophisticated address parser, has been integrated for more accurate analysis of inputs at
/v1/search. - Street names containing post-directionals (e.g. -
186 Tuskegee St SE Atlanta GA->186 Tuskegee St SouthEast Atlanta GA) are now treated the same as their pre-directional brethren. - 10/10, would release again - geocoding fallback rules that favor traditional geocoding behavior instead of search engine behavior
Another data-only release. Stay tuned for next week!
- Get excited for the addition of ✨ STREETS ✨! That's right, with this release Mapzen Search gets a brand new
streetlayer, which contains OSM street centroids. With this addition, if we can't find the exact address you're looking for we'll return the street record. Stay tuned for an in-depth blog post in the next few days. 👏
- Crikey! We noticed we weren't handling Australian province abbreviations, so we added support for them in our labels.
- Geonames ADM3 records now are correctly listed as localadmins, not venues.
- Our wonderful, now departed intern made sure Germanic street names are consistently handled (previously, some would end in -strasse while others ended in the abbreviation -str).
- Records with a Who's on First dependency now have that dependency listed in API responses.
No changes in functionality at all, just the freshest data! We did clean up some tests and do other work only visible to developers and those who run their own Pelias instance, but nothing major.
Stay tuned for next week's release where we already have some nice changes queued up.
- After much feedback we've added the
boundary.countryparameter for autocomplete! It works just like the one on the search endpoint. - To help make Leaflet maps display results better, we now use use the
lbl:bboxproperty on Who's on First records. This is useful for places like San Francisco where the mathematical bounding box is bigger than people expect. - The API was incorrectly warning against using the
boundary.circleparameter. Now it doesn't complain! - We've added a new
/v1/nearbyendpoint that is currently in early alpha! There's no documentation, probably some bugs, and any part of the interface is still subject to change. - Finally, we're now running Node.js 4 in production, rather than Node.js 0.12. For those running their own Pelias instance, be aware that we'll be dropping support for Node.js 0.12 in September. At first, things may work on Node.js 0.12, but we're very excited to finally start using ES2016, so that won't last too long.
Incremental release resolving the final outstanding tasks in the Elasticsearch 2 upgrade.
We have registered a new website http://pelias.io which has information about the milstones we have planned for the current quarter.
- Elasticsearch 2+ does not support co-ordinate wrapping as it did prior to the 2 release. Some front-ends allow users to 'wrap' around the globe. Floats outside of the normal -90/+90 -180/+180 geographic coordinate ranges cause Elasticsearch to error. We added a function to the API which unwraps these coordinates; providing better compatibility with these tools.
- We added
boroughas a possible layer for Geonames - Since the beginning of the project the Elasticsearch
_indexname has always been hard-coded as 'pelias', the index names configurable PRs allow this behaviour to be adjusted in your individual pelias config files. - We removed the focus.viewport API which was undocumented and never used outside of test suites.
Another bigger than usual release, we had some ops related challenges to resolve after the update to Elasticsearch 2, as well as some data issues, but we also have some great improvements in store!
- We use more of the population data in Who's on First, which really helps more relevant cities come up in searches.
- Searching for only records in certain layers in Geonames now works! We keep adding better handling of Geonames data but sometimes our API code doesn't keep up with those changes.
- Labels now include county names if there's no city (locality) info present. This helps with addresses that are outside the bounds of any city
- Capitalization across all OpenAddresess records is now more consistent. We've tried to properly capitalize all records that were either in all caps or all lowercase. This is better in general, although there are certainly exceptions, and we welcome bug reports for those cases.
- Geonames records for New York City boroughs like Manhattan and Brooklyn are now in the
borough, rather thanlocalitylayer. This makes them consistent with the records from Who's on First, which have been boroughs for some time. - Addresses in the Czech Republic now show the street name before the house number, in keeping with local customs
- When using the
/v1/placeendpoint, the source name can either be the full name or the abbreviation (like thesourcesparameter to the search and autocomplete endpoints). We love saving people some typing :) - We've made lots of internal changes like reducing the size of our documents, using a cleaner method to construct layer filter queries, removing dependencies on packages we've deprecated, and allowing the Elasticsearch index name to be configured for both the API and schema packages.
- In related internal changes news, we've also worked to make sure that all our code works with Node.js version 6, which was recently released! Support for Node.js 0.10, which is quite old and near end-of-life, is also starting to be removed.
We also have two known issues in this build:
- Some OpenAddresses records for the statewide data in Massachusetts, USA are incorrect. This is because of an issue when changing data sources that will be resolved in the next OpenAddresses build
- Geonames
localadminrecords, like the City of New York will have extra components in the label (in this case, "Brooklyn, New York"). The fix for this is merged but was accidentally omitted from this build. Look forward to it next week!
- Big news: We've finally upgraded to Elasticsearch 2.3! This brings improved performance and more importantly sets us up for lots of improvements from the new features of Elasticsearch 2. Elasticsearch 1.7 is no longer supported.
- As part of the Elasticsearch 2 upgrade we've also improved a few edge cases for searching for numeric values, and with single character tokens. You can read more in the Github issue for the upgrade.
- We've also fixed some lingering issues where a few places in Denmark were listed as being part of Sweden. This was due to the same data bug as mentioned in our recent blog post.
- The OpenAddresses importer now has better whitespace cleanup, so there won't be any extra spaces in street names.
- We recently added data to new layers in Geonames, but the API didn't know about it, and prevented you from searching for them. We fixed it.
- Who's on First importer: records now use the label centroid if it's present. The previous behavior was to always use the center of the record's bounding box. In cases like San Francisco, this caused the record to not show up where people expect!
- Openstreetmap importer: A bug in config parameter handling that caused admin lookup to be disabled when it shouldn't was fixed. Thanks to @dylanFrese for helping us catch this tricky one.
- We did it... we removed an Elasticsearch analyzer that was presumptuously assuming all queries were in English! The
k-stemminganalysis would do strange things like turn Daly into Dale, so finding "Daly City" was a challange. Well, no more! Word of warning, in/searchwe are now less forgiving when someone uses a plural version of a word where the real name is singular.
- All the extra 0's have been eradicated in addresses coming from OpenAddresses. You should not see any house numbers that reduce to 0 or any leading 0's in house numbers.
- Added the mysteriously missing
source_idproperty to response features. This property represents the original id at the source, if one existed, like in OSM and WOF. Where it didn't we made one up to help uniquely identify each record.
- Cleaned up some invalid address data from our OpenAddresses import by removing anything with words like
NULL,UNDEFINED, andUNAVAILABLE. - Improved error reporting in the API so users can decipher what went wrong much easier. More specifically, there are errors that Elasticsearch reports and we propogate up to the API response.
- A big improvements to autocomplete results coming from numerous bug fixes and improvements! More details can be found in the pull requests: pelias/schema#127 and pelias/api#526. Some highlights include:
- Single digit housenumbers like
8 Main Stcan be found more easily - Support for searching for the street name before the house number, as is common in many European countries, is improved.
- Searches that end in common words no longer return no results. These were being treated as stopwords internally in Elasticsearch. Now queries such as
Moscone Westwill work better
- Address parsing now works without spaces after commas. This was our bad. Feel free to leave those spaces out as long as you provide commas to delimit admin parts.
- Further streamlining of labels. You can expect the labels to a have more consistent and minimal feel. If the results are coming from New York, expect boroughs such as Manhattan, Brooklyn, Queens, etc. to be part of the label. You're welcome New Yorkers! ❤️
- Fixed a bug where specifying
layers=macrocountywould fail due to a typo in the API code. You can see how easy it is to mistypemacrocountyand not notice thatmacrocountryis incorrect. #onlyhuman
This release marks the official integration of the Mapzen Who's on First data set into Mapzen Search. This data is replacing Quattroshapes across the entire service. Any forward usage or references to Quattroshapes will be replaced with WhosOnFirst. This substitution allows us to fix long-standing encoding issues in administrative hierarchy place-names. We've also added a bounding box for individual features in the results, not only the all-encompassing bounding box at the top level of the geojson results. Also, the all-encompassing bounding box will extend to include the bounding boxes of all the features in the results, not only their centroids.
Another major improvement that many have been waiting for is the addition of more filters for the /autocomplete endpoint. Users can now ask /autocomplete to filter by layers and sources, as documented here.
See the detailed list of changes below for more specifics.
- Switched from
QuattroshapestoWhosOnFirstas the canonical source for administrative hierarchies and corresponding geometries. - No longer importing
Quattroshapesdata sinceWhosOnFirstcontains all those records and more. Going forward, any use ofquattroshapesorqsin queries will resolve towhosonfirstorwofautomatically. - New
bboxproperty has been added to individual results, for which geometry was available in the original source. This does not affect POI and address data. - Drastic improvements have been made to the label generation logic.
idandgidformat has changed to make the ids more unique.- New id format resolves previously outstanding bugs related to
geonamesids being invalid for lookup via the/placeendpoint. - Additional place-types have been introduced, such as
macroregion,macrocounty, andborough. gidvalues have been added for each parent in the admin hierarchies of results./autocompletenow allows filtering bysourcesandlayers.- Fixed a bug that allowed
/autocompleteto accept thesizeparameter. The default and only size of/autocompleteresults is now10, as originally intended.