About Me • Software Engineer & Data Nerd since 1997 • Doing “media stuff” since 2012 • Principal @ AOL since 2014 • @ieatkillerbees • http://samanthaquinones.com
What is Elasticsearch? • Near real-time (documents are available for search quickly after being indexed) search engine powered by Lucene • Clustered for H/A and performance via federation with shards and replicas
What’s it Used For? • Logging (we use Elasticsearch to centralize traffic logs, exception logs, and audit logs) • Content management and search • Statistical analysis
Connecting to Elasticsearch • Via Java, there are two native clients which connect to an ES cluster on port 9300 • Most commonly, we access Elasticsearch via HTTP API
Index Madness • Index is an overloaded term. • As a verb, to index a document is store a document in an index. This is analogous to an SQL INSERT operation. • As a noun, an index is a collection of documents. • Fields within a document have inverted indexes, similar to how a column in an SQL table may have an index.
Relevancy • When searching (in query context), results are scored by a relevancy algorithm • Results are presented in order from highest to lowest score
Aggregations • Run statistical operations over your data • Also near real-time! • Complex aggregations are abstracted away behind simple interfaces— you don’t need to be a statistician
Under the Hood • Elasticsearch is designed from the ground-up to run in a distributed fashion. • Indices (collections of documents) are partitioned in to shards. • Shards can be stored on a single or multiple nodes. • Shards are balanced across the cluster to improve performance • Shards are replicated for redundancy and high availability
What is a Cluster? • One or more nodes (servers) that work together to… • serve a dataset that exceeds the capacity of a single server… • provide federated indexing (writes) and searching (reads)… • provide H/A through sharing and replication of data
What is an Index? • An index is logically a collection of documents, roughly analogous to a database in MySQL • An index is in reality a namespace that points to one or more physical shards which contain data • When indexing a document, if the specified index does not exist, it will be created automatically
What are Shards? • Low-level units that hold a slice of available data • A shard represents a single instance of lucene and is fully- functional, self-contained search engine • Shards are either primary or replicas and are assigned to nodes
What is Replication? • Shards can have replicas • Replicas primarily provide redundancy for when shards/nodes fail • Replicas should not be allocated on the same node as the shard it replicates
allocation > $ http GET ':9200/_cat/allocation?v' shards disk.used disk.avail disk.total disk.percent host 33 2.6gb 21.8gb 24.4gb 10 host1 33 3gb 21.4gb 24.4gb 12 host2 34 2.6gb 21.8gb 24.4gb 10 host3 Show how many shards are allocated per node, with disk utilization info
count > $ http GET ':9200/_cat/count?v' epoch timestamp count 1453790185 06:36:25 182763 > $ http GET ‘:9200/_cat/count/posts?v’ epoch timestamp count 1453790467 06:41:07 164169 > $ http GET ‘:9200/_cat/count/posts.public?v’ epoch timestamp count 1453790472 06:41:12 164169= Display a count of documents in the cluster, or a specific index
fielddata > $ http -b GET ':9200/_cat/fielddata?v' id host ip node total site_id published 7tjeJNY3TMajqRkmYsJyrA host1 10.97.183.146 node1 1.1mb 170.1kb 996.5kb __xrpsKAQW6yyCY8luLQdQ host2 10.97.180.138 node2 1.6mb 329.3kb 1.3mb bdoNNXHXRryj22YqjnqECw host3 10.97.181.190 node3 1.1mb 154.7kb 991.7kb Shows how much memory is allocated to fielddata (metadata used for sorts)
indices > $ http -b GET 'eventhandler-prod.elasticsearch.amppublish.aws.aol.com:9200/_cat/indices?v' health status index pri rep docs.count docs.deleted store.size pri.store.size green open posts_561729df8ce4e 5 1 468629 20905 4gb 2gb green open slideshows 5 1 3893 6 86mb 43mb
shards > $ http -b GET ':9200/_cat/shards?v' index shard prirep state docs store ip node posts_561729df8ce4e 2 r STARTED 94019 410.5mb 10.97.180.138 host1 posts_561729df8ce4e 2 p STARTED 94019 412.7mb 10.97.181.190 host2 posts_561729df8ce4e 0 p STARTED 93307 413.6mb 10.97.183.146 host3 posts_561729df8ce4e 0 r STARTED 93307 415mb 10.97.180.138 host1 posts_561729df8ce4e 3 p STARTED 94182 407.1mb 10.97.183.146 host2 posts_561729df8ce4e 3 r STARTED 94182 403.4mb 10.97.180.138 host1 posts_561729df8ce4e 1 r STARTED 94130 447.1mb 10.97.180.138 host1 posts_561729df8ce4e 1 p STARTED 94130 447mb 10.97.181.190 host2 posts_561729df8ce4e 4 r STARTED 93299 421.5mb 10.97.183.146 host3 posts_561729df8ce4e 4 p STARTED 93299 398.8mb 10.97.181.190 host2
Metadata • _index - The index in which a document resides • _type - The class of object that a document represents • _id - The document’s unique identifier. Auto-generated when not provided
Retrieving Documents curl -X GET "http://localhost:9200/test_document/test/1" curl -X HEAD “http://localhost:9200/test_document/test/1" curl -X HEAD "http://localhost:9200/test_document/test/2"
Bulk API • Perform many operations in a single request • Efficient batching of actions • Bulk queries take the form of a stream of single-line JSON objects that define actions and document bodies
Bulk Actions • create - Index a document IFF it doesn’t exist already • index - Index a document, replacing it if it exists • update - Apply a partial update to a document • delete - Delete a document
Sizing Bulk Requests • Balance quantity of documents with size of documents • Docs list the sweet-spot between 5-15 MB per request • AOL Analytics Cluster indexes 5000 documents per batch (approx 7MB)
Searching Documents • Structured queries - queries against concrete fields like “title” or “score” which return specific documents. • Full-text queries - queries that find documents which match a search query and return them sorted by relevance
Search Elements • Mappings - Defines how data in fields are interpreted • Analysis - How text is parsed and processed to make it searchable • Query DSL - Elasticsearch’s query language
About Queries • Leaf Queries - Searches for a value in a given field. These queries are standalone. Examples: match, range, term • Compound Queries - Combinations of leaf queries and other compound queries which combine operations together either logically (e.g. bool queries) or alter their behavior (e.g. score queries)
Pagination Concerns • Since searches are distributed across multiple shards, paged queries must be sorted at each shard, combined, and resorted • The cost of paging in distributed data sets can increase exponentially • It is a wise practice to set limits to how many pages of results can be returned
Full Text Queries • match - Basic term matching query • multi_match - Match which spans multiple fields • common_terms - Match query which preferences uncommon words • query_string - Match documents using a search “mini-dsl” • simple_query_string - A simpler version of query_string that never throws exceptions, suitable for exposing to users
Term Queries • term - Search for an exact value • terms - Search for an exact value in multiple fields • range - Find documents where a value is in a certain range • exists - Find documents that have any non-null value in a field • missing - Inversion of `exists` • prefix - Match terms that begin with a string • wildcard - Match terms with a wildcard • regexp - Match terms against a regular expression • fuzzy - Match terms with configurable fuzziness
Compound Queries • constant_score - Wraps a query in filter context, giving all results a constant score • bool - Combines multiple leaf queries with `must`, `should`, `must_not` and `filter` clauses • dis_max - Similar to bool, but creates a union of subquery results scoring each document with the maximum score of the query that produced it • function_score - Modifies the scores of documents returned by a query . Useful for altering the distribution of results based on recency, popularity, etc. • boosting - Takes a `positive` and `negative` query, returning the results of `positive` while reducing the scores of documents that also match `negative` • filtered - Combines a query clause in query context with one in filter context • limit - Perform the query over a limited number of documents in each shard
What are Mappings? • Similar to schemas, they define the types of data found in fields • Determines how individual fields are analyzed & stored • Sets the format of date fields • Sets rules for mapping dynamic fields
Mapping Types • Indices have one or more mapping types which group documents logically. • Types contain meta fields, which can be used to customize metadata like _index, _id, _type, and _source • Types can also list fields that have consistent structure across types.
Dynamic vs Explicit Mapping • Dynamic fields are not defined prior to indexing • Elasticsearch selects the most likely type for dynamic fields, based on configurable rules • Explicit fields are defined exactly prior to indexing • Types cannot accept data that is the wrong type for an explicit mapping
Shared Fields • Fields that are defined in multiple mapping types must be identical if: • They have the same name • Live in the same index • Map to the same field internally
Dynamic Mappings • Mappings are generated when a type is created, if no mapping was previously specified. • Elasticsearch is good at identifying fields much of the time, but it’s far from perfect! • Fields can contain basic data-types, but importantly, mappings optimize a field for either structured (exact) or full-text searching
Structured Data vs Full Text • Exact values contain exact strings which are not subject to natural language interpretation. • Full-text values must be interpreted in the context of natural language
Natural Language • “us” can be interpreted differently in natural language • Abbreviation for “United States” • The English dative personal pronoun • An alternative symbol for µs • The French word us
Analyzing Text • Elasticsearch is optimized for full text search • Text is analyzed in a two-step process • First, text is tokenized in to individual terms • Second, terms are normalized through a filter
Analyzers • Analyzers perform the analysis process • Character filters clean up text, removing or modifying the text • Tokenizers break the text down in to terms • Token filters modify, remove, or add terms
Standard Analyzer • General purpose analyzer that works for most natural language. • Splits text on word boundaries, removes punctuation, and lowercases all tokens.
Analyzers • Analyzers are applied when documents are indexed • Analyzers are applied when a full-text search is performed against a field, in order to produce the correct set of terms to search for
Character Filters • html_strip - Removes HTML from text • mapping - Filter based on a map of original → new ( { “ph”: “f” }) • pattern_replace - Similar to mapping, using regular expressions
Index Templates • Template mappings that are applied to newly created indices • Templates also contain index configuration information • Powerful when combined with dated indices
Scoring • Scoring is based on a boolean model and scoring function • Boolean model applies AND/OR logic to an inverse index to produce a list of matching documents
Term Frequency • Terms that appear frequently in a document increase the document’s relevancy score. • term_frequency(term in document) = √number_of_appearances
Inverse Document Frequency • Terms that appear in many documents reduce a document’s relevancy score • inverse_doc_frequency(term) = 1 + log(number_of_docs / (frequency + 1))
Example from the Docs • Given the text “quick brown fox” the term “fox” scores… • Term Frequency: 1.0 • Inverse Doc Frequency: 0.30685282 • Field Norm: 0.5 • Score: 0.15342641
Importing Energy Data curl -X PUT "http://localhost:9200/energy_use" --data-binary "@queries/ mapping_energy.json" curl -X PUT "http://localhost:9200/_bulk" --data-binary "@queries/ bulk_insert_energy_data.json" curl -X GET "http://localhost:9200/energy_use/_search"