This page describes the Wikimedia Foundation's activities surrounding our sites' search functionality. Our current project is to replace our legacy lsearchd system with a new system based on Elasticsearch (using a new extension called CirrusSearch). This project started in June 2013, with the migration slated to last until early 2014.
- 1 Status
- 2 Rationale
- 3 Timeline
- 4 Solr vs Elasticsearch
- 5 Elasticsearch implementation plan
- 6 Search Weighting Ideas
- 7 Items in search results
- 8 General Search Behaviors
- 9 Research
- 10 Documents
- 11 Links
Now that we're stable again I've started working on the problem at the root:
- We're getting more servers. We're about doubling the cluster size.
- I'm putting together more optimizations to the portion of Cirrus that fell over (working set). If everything goes as planned we'll reduce it by about 80%. They swap indexing performance for search performance.
I'm going to start rolling those optimization out on Monday July 28th. They won't go everywhere immediately, but I'll roll them in as see how the index time performance hit effects us.Also: these changes will change result relevance some. In my local testing it looked like everything still worked: title still beats redirect still beats category still beat heading still beats article lead in still beats text still beats image captions and file contents. BUT I still expect things to shift around a bit. Please let me know if you see anything fishy.
Rationale[edit | edit source]
The Wikimedia search infrastructure hasn't had significant development work for many years. The current system is based on homegrown layer (named "lsearchd") on top of Lucene. The problem lsearchd solves has since been tackled by much larger projects such as Solr and Elasticsearch. The lsearchd system frequently breaks in ways that are difficult to diagnose[clarification needed], and generally makes our Operations staff sad.
Goals for our current effort:
- Make our existing tools more robust
- Improve logging in our existing tools to make problems easier to diagnose
- Migrate away from lsearchd to Solr or something similar
Our current search infrastructure is highly outdated and difficult to manage due to tons of custom code. We are now replacing lsearchd with Elasticsearch (which is also a layer on Lucene), as it's very stable, contains many of the features we need, and doesn't require nearly as much custom code to support. What custom code we write will be incorporated in a MediaWiki extension called CirrusSearch.
Timeline[edit | edit source]
This page is a timeline for future deployments of Cirrus as primary backends to wikis. Our general goal is to deploy CirrusSearch (backed by Elasticsearch) as the primary search backend for all wikis by the end of
2013(ha!) September 2014.
Wikis[edit | edit source]
This table is of the current plan. We imagine the current plan to change frequently. Historically we've been pretty bad at keeping this up to date when the plan slips but we'll try to be better in the future.
All dates below are from before we slipped the schedule: we'll put a new one together the last week of July.
|About 30 (mostly) wikipedias|
|About 30 more (mostly) wikipedias|
Solr vs Elasticsearch[edit | edit source]
We spent some time looking at search systems we could use and it became pretty apparent that the thought leaders in the open source world for search are Solr and Elasticsearch. We spent a few weeks with each and decided to build on Elasticsearch because of its wonderful suggester, easily composable queries and good documentation. We are also happy with the process of submitting changes upstream to Elasticsearch.
Elasticsearch implementation plan[edit | edit source]
Core Search[edit | edit source]
We're working feverishly on CirrusSearch and we're deploying it to wikis on a volunteer basis. We're actively looking for new volunteers so if you want in email firstname.lastname@example.org.
For other WMF applications[edit | edit source]
GeoData[edit | edit source]
We've just started looking at how to move GeoData to Elasticsearch. For now, it'll remain in Solr with plans to migrate it to Elasticsearch when time permits. Some considerations:
The index is relatively small (so no need to make it distributed), but requires a lot of computational power to work with. Full-text search is not currently used. Currently, data from all the wikis is stored in the same core, in the future we will need to split data to many cores (the puppet changes for using multiple cores with shared configuration/schema are here, needs more work).
- Load expectations: unclear, but will be high if we start using it heavily e.g. for maps display.
- Backups: not really needed - if master is down just switch to a slave. If all servers are down, reindexing from scratch is quick.
- Note: because GeoData's schema is very stripped-down, /admin/ping doesn't work - should be remembered if someone wants to rewrite the current monitoring.
Translation Memory[edit | edit source]
- Niklas wants to work with Chad & Nik to figure out what is needed
- "This spring/summer"
Search Weighting Ideas[edit | edit source]
Some things that could be factored into search results, in rough proposed grouping order:
(+) positive impact on default ranking (-) negative impact on default ranking
Relationship[edit | edit source]
- In-article searches (when I'm reading an article, I want to search within it) (+)
- GeoLoc of user when searching (+)
- Articles nearness (geographically) to current article (+)
- Articles linked from current page (+)
- Wikidata related items (+)
- Categories on articles (+)
Relevance[edit | edit source]
- Article "meshiness" (number of articles that link to article, number of articles that article links to) (+)
- External search referrer terms saved into article metadata (not a thing yet) (+)
- Article importance (per wikiprojects) (+) (-)
- Recency of last edit (+) (-)
- Article recency (creation) (+)
- Matches in title, headings, body text, alternate text, references (weighted differently) (+) (-)
- Other wiki search results
- Is this a weighting thing, or a content type thing (e.g. if it has related results on other wikis it's ranked higher)?
Quality[edit | edit source]
- Quality of an article (i.e. featured) (+)
- Article quality (per wikiprojects) (+) (-)
- Content with article issue templates (-)
- Calls to action (number of article issues, missing images, etc.) (-)
- Stub status (-)
Aggregate User behavior[edit | edit source]
- Top searches in the last hour/day/week (+)
- Search terms that were followed vs unfollowed (+) (-)
- Recent pages arrived at via external search (+)
Items in search results[edit | edit source]
Existing[edit | edit source]
- Articles / Talk
- User / Talk
- File / Talk
- Category / Talk
- Lists / Talk
- Portal / Talk
- Help / Talk
- Template / Talk
- Module / Talk
- Wikipedia / Talk
- Education Program / Talk
- TimedText (captions/subtitles) / Talk
- MediaWiki / Talk
- Book / Talk
New/Modified[edit | edit source]
- Other media types(?)
- Article text
- Article sections
- Preference sections
- Tea house
- Reference desk
- Village Pump
Sister Project Results[edit | edit source]
- Wiktionary results
- Wikivoyage results (desktop only)
- Wikidata items
Contextual Search[edit | edit source]
Flow[edit | edit source]
General Flow search[edit | edit source]
- Topic titles
- Board titles (user names/preferred names)
- Full text search
- Board Descriptions
User Mention search[edit | edit source]
- Users participating in the current topic
- Users mentioned in the current topic
- Users participating on current board
- Users mentioned in current board
- Users mentioned my me in the last X days
- Users whom I follow (Gasp! This doesn't exist yet!)
Category/Tags[edit | edit source]
- Recently applied tags
Template/Transclusion Search[edit | edit source]
- Recently Used
- Most common templates inserted from current context
- Most common template inserted after last inserted
General Search Behaviors[edit | edit source]
- When zero results are found for exact match but "did you mean" is shown, show special header with did you mean note and create article note, but show did you mean results instead.
- In full text search show exact text page title matches in alternate visual appearance
Research[edit | edit source]
- Show floating right side microsurvey widget to rate search results
- Allow users to up and down vote results as relevant to inform weighting hypothesis
- Analyze search/followed link pair
- Run microsurvey "is this what you were looking for" on results page?
- Can we determine if it was what they were looking for without asking?
- amount of page viewed
- back to results rather than following link on page or bounce
Documents[edit | edit source]
- Search documentation on Wikitech: wikitech:Search
- Ram's setup instructions: wikitech:User:Ram/Search
- Some notes from Brion in 2008
- The MWSearch extension provides a SearchEngine subclass which contacts Wikimedia's Lucene-based search server. This replaces the older LuceneSearch extension which reimplemented the entire Special:Search page.
- /2013-02 discussion - Discussion with Rainman about how the current system works
- Video of May 2014 Wikimedia tech talk about ElasticSearch
Links[edit | edit source]
- Testing URLs: &srbackend=CirrusSearch, &srbackend=LuceneSearch