Dailymotion Elasticsearch June 10, 2014 Meetup Elasticsearch France #7 Cédric Hourcade Core developer at Dailymotion twitter.com/hced

> video search @ Dailymotion 1 > cluster overview and indexation 2 > query and score 3 > sharding 4 > benchmarks, tuning 5 > questions?

elasticsearch cluster cluster and indexation

search cluster cluster and indexation > 10 nodes for video search _ one main video index _ 5 shards with 1 replica _ nodes : 32 cores, 48 gb RAM, 15k disks > 600 to 1000 search requests per second > end-to-end response time < 40 ms

search cluster cluster and indexation elasticsearch cluster 10 nodes mysql farm index constantly

search cluster cluster and indexation elasticsearch cluster 10 nodes mysql farm elasticsearch indexer 2 nodes index constantly filter data for search update if hash changed hash data

search query and score

"query" : { "function_score": { "query": { "custom_common": { … } }, "script_score": { "script": "custom_scorer", "lang": "native" }, "scoring_score": "multiply" } } search query and score > custom query x > custom scorer = > score

search query and score Scorer > custom scorer _ only slightly alter the query score _ take into account: recency, popularity, etc. > boosted filters and scripts when testing > native java for performance

search query and score Query > we need to keep control of the query base score > problem is our text content is thin _ short title, a few tags _ a more or less relevant description > bare bones TF-IDF may not be suitable _ TF not that relevant to us

search query and score > BM25: reduce importance of document length > why common terms query _ increase performance _ ignore popular terms when searching _ but still use them for scoring _ like a real time specialized stop words list similarity: my_bm25: type: BM25 b: 0.001

> ignore inexistent terms in query > boost repeated terms (TF) only if repeated in query a doc titled “A A A game” has a better score than “A game” only when explicitly searching for “A A A” > boost term by position in query and documents search query and score brown fox zerzer brown fox zerzer the quick brown fox jumps. ^1.1 ^1.07 ^1.05 ^1.03 ^1.02

> keep both stemmed and original terms > score them with dis_max (tie_breaker = 0) > disable coord factor for consistent scoring I like dogs i like dogs dog token 1 2 3 position search query and score "field": "dogs" "dis_max": { "tie_breaker" : 0, "queries" : [ { "term": { "field": "dogs" } }, { "term": { "field": "dog" } } ...

sharding what suits us

sharding what suits us > less shards make a query slower > but not 16 times slower (112 ms vs 12 ms) index / 16 shards index / 1 shard 1 ms request handling 1 ms request handling 10 ms shard 0 (9ms) shard 1 (10ms) shard 2 (6ms) shard 3 (9ms) 110 ms shard 0 shard 4 (10ms) shard 5 (10ms) shard 6 (9ms) shard 7 (10ms) shard 8 (10ms) shard 9 (8ms) shard 10 (5ms) shard 11 (10ms) shard 12 (7ms) shard 13 (7ms) shard 14 (10ms) shard 15 (10ms) 1 ms return result 1 ms return result 12 ms 112 ms

sharding what suits us > takes more resources > everything runs at 100 % for each query > less requests per second for the same hardware 9 ms + 10 ms + 6 ms + (…) = 140 ms shard 0 (9ms) shard 1 (10ms) shard 2 (6ms) shard 3 (9ms) 110 ms shard 0 shard 4 (10ms) shard 5 (10ms) shard 6 (9ms) shard 7 (10ms) shard 8 (10ms) shard 9 (8ms) shard 10 (5ms) shard 11 (10ms) shard 12 (7ms) shard 13 (7ms) shard 14 (10ms) shard 15 (10ms) 140 ms spent by the shards 110 ms spent

sharding what suits us Before > we used to have 40 shards on 18 nodes _ ~2 millions docs per shard _ 3 gb by shards _ ~ 120 gb total index size > cluster was very loaded _ every single query was hitting all the nodes _ response times could have been better

sharding what suits us After > we now have 5 shards on 10 nodes > cluster run smoother, less load _ only 5 nodes involved per query _ it handles many times more requests

sharding what suits us less data! _ ~10 millions docs per shard _ 4 gb by shards _ ~ 25 gb total index size > only data we need right now _ { "_source" : false } _ round numbers and dates _ { "precision_step" : 2147483647 } > less updates, faster indexation, rebalance, merges...

sharding what suits us drawbacks > queries taken individually are slower… > but only marginally slower _ eg: 7 ms instead of 5 ms > but some slower queries became more noticeable

how do we test benchmarks, tuning

how do we test benchmarks, tuning load test _ benchmark with Tsung _ dedicated test cluster _ run real queries, lots of them _ aim for our expected load _ monitor everything _ reshard, change schema _ set masters, data-only nodes... repeat

how do we test benchmarks, tuning use warmers > warm segments after each merge _ prevent slow first queries > set it up to build cache for the filters we use > zero reasons for not using them

how do we test benchmarks, tuning "constant_score": { "filter": { "term": { "visible": "yes" } } } "constant_score": { "filter": { "bool": { "_cache": true, "must": [ { "term": { "visible": "yes" } }, { "range": { "age": { "from": 18, "to": 30 } } } ] ...

how do we test benchmarks, tuning query testing > to test a particular query raw performance _ one index, one shard _ millions of simple documents _ merged in one segment _ with some deletes

? how do we test benchmarks, tuning { "query": { "filtered": { "query": { "match": { "title": "some very popular terms" } }, "filter": { "term": { "user": "cedric" } } } } }

! how do we test benchmarks, tuning { "query": { "filtered": { "strategy": "leap_frog", "query": { "match": { "title": "some very popular terms" } }, "filter": { "term": { "user": "cedric" } } } } }

how do we test benchmarks, tuning > we also use Elasticsearch to just filter and sort > these queries match millions of documents _ they are slow _ even when terms are cached _ iterating, scoring and sorting is tedious

how do we test benchmarks, tuning query "sort": { "created": "desc" }, "query": { "bool": { "must": [ { "term": { "public": true } } ] } }

how do we test benchmarks, tuning query result "sort": { "created": "desc" }, "query": { "bool": { "must": [ { "term": { "public": true } } ] } } "took": 695 "hits": { "total": 79582599 }

how do we test benchmarks, tuning > we know our data > we can help our query

how do we test benchmarks, tuning query "sort": { "created": "desc" }, "query": { "bool": { "must": [ { "term": { "public": true } }, { "range": { "created": { "from": "2014-06-03" } } } ] ... > with a range filter on the sorted field

how do we test benchmarks, tuning query result "sort": { "created": "desc" }, "query": { "bool": { "must": [ { "term": { "public": true } }, { "range": { "created": { "from": "2014-06-03" } } } ] ... "took": 15 "hits": { "total": 92312 } // Same top docs // returned

how do we test benchmarks, tuning > what if there are not enough hits? _ re-run the query without the filter > we use a custom query to do just that! _ breaks once it matches enough hits _ runs at segment level _ no round-trips

"sort": { "created": "desc" }, "query": { "break_once": { "minimum_hits": 100, "query": { { "term": { "public": true } }, "filters": [ { "range": { "created": { "from": "2014-06-03" } } }, { "range": { "created": { "from": "2014-05-03” } } }, { "range": { "created": { "from": "2014-01-03" } } } ] } } how do we test benchmarks, tuning stop once there are enough hits

No content

thank you > index only what you need now > shard today, reshard tomorrow > benchmark to find what suits you best > test and optimize your queries

thank you > questions?