Profile API

Profile API

The Profile API is a debugging tool and adds significant overhead to search execution.

Provides detailed timing information about the execution of individual components in a search request.

Description

The Profile API gives the user insight into how search requests are executed at a low level so that the user can understand why certain requests are slow, and take steps to improve them. Note that the Profile API, amongst other things, doesn’t measure network latency, time spent while the requests spends in queues, or while merging shard responses on the coordinating node.

The output from the Profile API is very verbose, especially for complicated requests executed across many shards. Pretty-printing the response is recommended to help understand the output.

Examples

Any _search request can be profiled by adding a top-level profile parameter:

  1. GET /my-index-000001/_search
  2. {
  3.   "profile": true,
  4.   "query" : {
  5.     "match" : { "message" : "GET /search" }
  6.   }
  7. }

Setting the top-level profile parameter to true will enable profiling for the search.

The API returns the following result:

  1. {
  2.   "took": 25,
  3.   "timed_out": false,
  4.   "_shards": {
  5.     "total": 1,
  6.     "successful": 1,
  7.     "skipped": 0,
  8.     "failed": 0
  9.   },
  10.   "hits": {
  11.     "total": {
  12.       "value": 5,
  13.       "relation": "eq"
  14.     },
  15.     "max_score": 0.17402273,
  16.     "hits": [...] 
  17.   },
  18.   "profile": {
  19.     "shards": [
  20.       {
  21.         "id": "[2aE02wS1R8q_QFnYu6vDVQ][my-index-000001][0]",
  22.         "searches": [
  23.           {
  24.             "query": [
  25.               {
  26.                 "type": "BooleanQuery",
  27.                 "description": "message:get message:search",
  28.                 "time_in_nanos" : 11972972,
  29.                 "breakdown" : {
  30.                   "set_min_competitive_score_count": 0,
  31.                   "match_count": 5,
  32.                   "shallow_advance_count": 0,
  33.                   "set_min_competitive_score": 0,
  34.                   "next_doc": 39022,
  35.                   "match": 4456,
  36.                   "next_doc_count": 5,
  37.                   "score_count": 5,
  38.                   "compute_max_score_count": 0,
  39.                   "compute_max_score": 0,
  40.                   "advance": 84525,
  41.                   "advance_count": 1,
  42.                   "score": 37779,
  43.                   "build_scorer_count": 2,
  44.                   "create_weight": 4694895,
  45.                   "shallow_advance": 0,
  46.                   "create_weight_count": 1,
  47.                   "build_scorer": 7112295
  48.                 },
  49.                 "children": [
  50.                   {
  51.                     "type": "TermQuery",
  52.                     "description": "message:get",
  53.                     "time_in_nanos": 3801935,
  54.                     "breakdown": {
  55.                       "set_min_competitive_score_count": 0,
  56.                       "match_count": 0,
  57.                       "shallow_advance_count": 3,
  58.                       "set_min_competitive_score": 0,
  59.                       "next_doc": 0,
  60.                       "match": 0,
  61.                       "next_doc_count": 0,
  62.                       "score_count": 5,
  63.                       "compute_max_score_count": 3,
  64.                       "compute_max_score": 32487,
  65.                       "advance": 5749,
  66.                       "advance_count": 6,
  67.                       "score": 16219,
  68.                       "build_scorer_count": 3,
  69.                       "create_weight": 2382719,
  70.                       "shallow_advance": 9754,
  71.                       "create_weight_count": 1,
  72.                       "build_scorer": 1355007
  73.                     }
  74.                   },
  75.                   {
  76.                     "type": "TermQuery",
  77.                     "description": "message:search",
  78.                     "time_in_nanos": 205654,
  79.                     "breakdown": {
  80.                       "set_min_competitive_score_count": 0,
  81.                       "match_count": 0,
  82.                       "shallow_advance_count": 3,
  83.                       "set_min_competitive_score": 0,
  84.                       "next_doc": 0,
  85.                       "match": 0,
  86.                       "next_doc_count": 0,
  87.                       "score_count": 5,
  88.                       "compute_max_score_count": 3,
  89.                       "compute_max_score": 6678,
  90.                       "advance": 12733,
  91.                       "advance_count": 6,
  92.                       "score": 6627,
  93.                       "build_scorer_count": 3,
  94.                       "create_weight": 130951,
  95.                       "shallow_advance": 2512,
  96.                       "create_weight_count": 1,
  97.                       "build_scorer": 46153
  98.                     }
  99.                   }
  100.                 ]
  101.               }
  102.             ],
  103.             "rewrite_time": 451233,
  104.             "collector": [
  105.               {
  106.                 "name": "SimpleTopScoreDocCollector",
  107.                 "reason": "search_top_hits",
  108.                 "time_in_nanos": 775274
  109.               }
  110.             ]
  111.           }
  112.         ],
  113.         "aggregations": [],
  114.         "fetch": {
  115.           "type": "fetch",
  116.           "description": "",
  117.           "time_in_nanos": 660555,
  118.           "breakdown": {
  119.             "next_reader": 7292,
  120.             "next_reader_count": 1,
  121.             "load_stored_fields": 299325,
  122.             "load_stored_fields_count": 5
  123.           },
  124.           "debug": {
  125.             "stored_fields": ["_id", "_routing", "_source"]
  126.           },
  127.           "children": [
  128.             {
  129.               "type": "FetchSourcePhase",
  130.               "description": "",
  131.               "time_in_nanos": 20443,
  132.               "breakdown": {
  133.                 "next_reader": 745,
  134.                 "next_reader_count": 1,
  135.                 "process": 19698,
  136.                 "process_count": 5
  137.               },
  138.               "debug": {
  139.                 "fast_path": 5
  140.               }
  141.             }
  142.           ]
  143.         }
  144.       }
  145.     ]
  146.   }
  147. }

Search results are returned, but were omitted here for brevity.

Even for a simple query, the response is relatively complicated. Let’s break it down piece-by-piece before moving to more complex examples.

The overall structure of the profile response is as follows:

  1. {
  2.    "profile": {
  3.         "shards": [
  4.            {
  5.               "id": "[2aE02wS1R8q_QFnYu6vDVQ][my-index-000001][0]",  
  6.               "searches": [
  7.                  {
  8.                     "query": [...],             
  9.                     "rewrite_time": 51443,      
  10.                     "collector": [...]          
  11.                  }
  12.               ],
  13.               "aggregations": [...],            
  14.               "fetch": {...}                    
  15.            }
  16.         ]
  17.      }
  18. }

A profile is returned for each shard that participated in the response, and is identified by a unique ID.

Query timings and other debugging information.

The cumulative rewrite time.

Names and invocation timings for each collector.

Aggregation timings, invocation counts, and debug information.

Fetch timing and debug information.

Because a search request may be executed against one or more shards in an index, and a search may cover one or more indices, the top level element in the profile response is an array of shard objects. Each shard object lists its id which uniquely identifies the shard. The ID’s format is [nodeID][indexName][shardID].

The profile itself may consist of one or more “searches”, where a search is a query executed against the underlying Lucene index. Most search requests submitted by the user will only execute a single search against the Lucene index. But occasionally multiple searches will be executed, such as including a global aggregation (which needs to execute a secondary “match_all” query for the global context).

Inside each search object there will be two arrays of profiled information: a query array and a collector array. Alongside the search object is an aggregations object that contains the profile information for the aggregations. In the future, more sections may be added, such as suggest, highlight, etc.

There will also be a rewrite metric showing the total time spent rewriting the query (in nanoseconds).

As with other statistics apis, the Profile API supports human readable outputs. This can be turned on by adding ?human=true to the query string. In this case, the output contains the additional time field containing rounded, human readable timing information (e.g. "time": "391,9ms", "time": "123.3micros").

Profiling Queries

The details provided by the Profile API directly expose Lucene class names and concepts, which means that complete interpretation of the results require fairly advanced knowledge of Lucene. This page attempts to give a crash-course in how Lucene executes queries so that you can use the Profile API to successfully diagnose and debug queries, but it is only an overview. For complete understanding, please refer to Lucene’s documentation and, in places, the code.

With that said, a complete understanding is often not required to fix a slow query. It is usually sufficient to see that a particular component of a query is slow, and not necessarily understand why the advance phase of that query is the cause, for example.

query Section

The query section contains detailed timing of the query tree executed by Lucene on a particular shard. The overall structure of this query tree will resemble your original Elasticsearch query, but may be slightly (or sometimes very) different. It will also use similar but not always identical naming. Using our previous match query example, let’s analyze the query section:

  1. "query": [
  2.     {
  3.        "type": "BooleanQuery",
  4.        "description": "message:get message:search",
  5.        "time_in_nanos": "11972972",
  6.        "breakdown": {...},               
  7.        "children": [
  8.           {
  9.              "type": "TermQuery",
  10.              "description": "message:get",
  11.              "time_in_nanos": "3801935",
  12.              "breakdown": {...}
  13.           },
  14.           {
  15.              "type": "TermQuery",
  16.              "description": "message:search",
  17.              "time_in_nanos": "205654",
  18.              "breakdown": {...}
  19.           }
  20.        ]
  21.     }
  22. ]

The breakdown timings are omitted for simplicity.

Based on the profile structure, we can see that our match query was rewritten by Lucene into a BooleanQuery with two clauses (both holding a TermQuery). The type field displays the Lucene class name, and often aligns with the equivalent name in Elasticsearch. The description field displays the Lucene explanation text for the query, and is made available to help differentiating between parts of your query (e.g. both message:get and message:search are TermQuery’s and would appear identical otherwise.

The time_in_nanos field shows that this query took ~11.9ms for the entire BooleanQuery to execute. The recorded time is inclusive of all children.

The breakdown field will give detailed stats about how the time was spent, we’ll look at that in a moment. Finally, the children array lists any sub-queries that may be present. Because we searched for two values (“get search”), our BooleanQuery holds two children TermQueries. They have identical information (type, time, breakdown, etc). Children are allowed to have their own children.

Timing Breakdown

The breakdown component lists detailed timing statistics about low-level Lucene execution:

  1. "breakdown": {
  2.   "set_min_competitive_score_count": 0,
  3.   "match_count": 5,
  4.   "shallow_advance_count": 0,
  5.   "set_min_competitive_score": 0,
  6.   "next_doc": 39022,
  7.   "match": 4456,
  8.   "next_doc_count": 5,
  9.   "score_count": 5,
  10.   "compute_max_score_count": 0,
  11.   "compute_max_score": 0,
  12.   "advance": 84525,
  13.   "advance_count": 1,
  14.   "score": 37779,
  15.   "build_scorer_count": 2,
  16.   "create_weight": 4694895,
  17.   "shallow_advance": 0,
  18.   "create_weight_count": 1,
  19.   "build_scorer": 7112295
  20. }

Timings are listed in wall-clock nanoseconds and are not normalized at all. All caveats about the overall time_in_nanos apply here. The intention of the breakdown is to give you a feel for A) what machinery in Lucene is actually eating time, and B) the magnitude of differences in times between the various components. Like the overall time, the breakdown is inclusive of all children times.

The meaning of the stats are as follows:

All parameters:

create_weight

A Query in Lucene must be capable of reuse across multiple IndexSearchers (think of it as the engine that executes a search against a specific Lucene Index). This puts Lucene in a tricky spot, since many queries need to accumulate temporary state/statistics associated with the index it is being used against, but the Query contract mandates that it must be immutable.

To get around this, Lucene asks each query to generate a Weight object which acts as a temporary context object to hold state associated with this particular (IndexSearcher, Query) tuple. The weight metric shows how long this process takes

build_scorer

This parameter shows how long it takes to build a Scorer for the query. A Scorer is the mechanism that iterates over matching documents and generates a score per-document (e.g. how well does “foo” match the document?). Note, this records the time required to generate the Scorer object, not actually score the documents. Some queries have faster or slower initialization of the Scorer, depending on optimizations, complexity, etc.

This may also show timing associated with caching, if enabled and/or applicable for the query

next_doc

The Lucene method next_doc returns Doc ID of the next document matching the query. This statistic shows the time it takes to determine which document is the next match, a process that varies considerably depending on the nature of the query. Next_doc is a specialized form of advance() which is more convenient for many queries in Lucene. It is equivalent to advance(docId() + 1)

advance

advance is the “lower level” version of next_doc: it serves the same purpose of finding the next matching doc, but requires the calling query to perform extra tasks such as identifying and moving past skips, etc. However, not all queries can use next_doc, so advance is also timed for those queries.

Conjunctions (e.g. must clauses in a Boolean) are typical consumers of advance

match

Some queries, such as phrase queries, match documents using a “two-phase” process. First, the document is “approximately” matched, and if it matches approximately, it is checked a second time with a more rigorous (and expensive) process. The second phase verification is what the match statistic measures.

For example, a phrase query first checks a document approximately by ensuring all terms in the phrase are present in the doc. If all the terms are present, it then executes the second phase verification to ensure the terms are in-order to form the phrase, which is relatively more expensive than just checking for presence of the terms.

Because this two-phase process is only used by a handful of queries, the match statistic is often zero

score

This records the time taken to score a particular document via its Scorer

*_count

Records the number of invocations of the particular method. For example, “next_doc_count”: 2, means the nextDoc() method was called on two different documents. This can be used to help judge how selective queries are, by comparing counts between different query components.

collectors Section

The Collectors portion of the response shows high-level execution details. Lucene works by defining a “Collector” which is responsible for coordinating the traversal, scoring, and collection of matching documents. Collectors are also how a single query can record aggregation results, execute unscoped “global” queries, execute post-query filters, etc.

Looking at the previous example:

  1. "collector": [
  2.   {
  3.     "name": "SimpleTopScoreDocCollector",
  4.     "reason": "search_top_hits",
  5.     "time_in_nanos": 775274
  6.   }
  7. ]

We see a single collector named SimpleTopScoreDocCollector wrapped into CancellableCollector. SimpleTopScoreDocCollector is the default “scoring and sorting” Collector used by Elasticsearch. The reason field attempts to give a plain English description of the class name. The time_in_nanos is similar to the time in the Query tree: a wall-clock time inclusive of all children. Similarly, children lists all sub-collectors. The CancellableCollector that wraps SimpleTopScoreDocCollector is used by Elasticsearch to detect if the current search was cancelled and stop collecting documents as soon as it occurs.

It should be noted that Collector times are independent from the Query times. They are calculated, combined, and normalized independently! Due to the nature of Lucene’s execution, it is impossible to “merge” the times from the Collectors into the Query section, so they are displayed in separate portions.

For reference, the various collector reasons are:

search_sorted

A collector that scores and sorts documents. This is the most common collector and will be seen in most simple searches

search_count

A collector that only counts the number of documents that match the query, but does not fetch the source. This is seen when size: 0 is specified

search_terminate_after_count

A collector that terminates search execution after n matching documents have been found. This is seen when the terminate_after_count query parameter has been specified

search_min_score

A collector that only returns matching documents that have a score greater than n. This is seen when the top-level parameter min_score has been specified.

search_multi

A collector that wraps several other collectors. This is seen when combinations of search, aggregations, global aggs, and post_filters are combined in a single search.

search_timeout

A collector that halts execution after a specified period of time. This is seen when a timeout top-level parameter has been specified.

aggregation

A collector that Elasticsearch uses to run aggregations against the query scope. A single aggregation collector is used to collect documents for all aggregations, so you will see a list of aggregations in the name rather.

global_aggregation

A collector that executes an aggregation against the global query scope, rather than the specified query. Because the global scope is necessarily different from the executed query, it must execute its own match_all query (which you will see added to the Query section) to collect your entire dataset

rewrite Section

All queries in Lucene undergo a “rewriting” process. A query (and its sub-queries) may be rewritten one or more times, and the process continues until the query stops changing. This process allows Lucene to perform optimizations, such as removing redundant clauses, replacing one query for a more efficient execution path, etc. For example a Boolean → Boolean → TermQuery can be rewritten to a TermQuery, because all the Booleans are unnecessary in this case.

The rewriting process is complex and difficult to display, since queries can change drastically. Rather than showing the intermediate results, the total rewrite time is simply displayed as a value (in nanoseconds). This value is cumulative and contains the total time for all queries being rewritten.

A more complex example

To demonstrate a slightly more complex query and the associated results, we can profile the following query:

  1. GET /my-index-000001/_search
  2. {
  3.   "profile": true,
  4.   "query": {
  5.     "term": {
  6.       "user.id": {
  7.         "value": "elkbee"
  8.       }
  9.     }
  10.   },
  11.   "aggs": {
  12.     "my_scoped_agg": {
  13.       "terms": {
  14.         "field": "http.response.status_code"
  15.       }
  16.     },
  17.     "my_global_agg": {
  18.       "global": {},
  19.       "aggs": {
  20.         "my_level_agg": {
  21.           "terms": {
  22.             "field": "http.response.status_code"
  23.           }
  24.         }
  25.       }
  26.     }
  27.   },
  28.   "post_filter": {
  29.     "match": {
  30.       "message": "search"
  31.     }
  32.   }
  33. }

This example has:

  • A query
  • A scoped aggregation
  • A global aggregation
  • A post_filter

The API returns the following result:

  1. {
  2.   ...
  3.   "profile": {
  4.     "shards": [
  5.       {
  6.         "id": "[P6-vulHtQRWuD4YnubWb7A][my-index-000001][0]",
  7.         "searches": [
  8.           {
  9.             "query": [
  10.               {
  11.                 "type": "TermQuery",
  12.                 "description": "message:search",
  13.                 "time_in_nanos": 141618,
  14.                 "breakdown": {
  15.                   "set_min_competitive_score_count": 0,
  16.                   "match_count": 0,
  17.                   "shallow_advance_count": 0,
  18.                   "set_min_competitive_score": 0,
  19.                   "next_doc": 0,
  20.                   "match": 0,
  21.                   "next_doc_count": 0,
  22.                   "score_count": 0,
  23.                   "compute_max_score_count": 0,
  24.                   "compute_max_score": 0,
  25.                   "advance": 3942,
  26.                   "advance_count": 4,
  27.                   "score": 0,
  28.                   "build_scorer_count": 2,
  29.                   "create_weight": 38380,
  30.                   "shallow_advance": 0,
  31.                   "create_weight_count": 1,
  32.                   "build_scorer": 99296
  33.                 }
  34.               },
  35.               {
  36.                 "type": "TermQuery",
  37.                 "description": "user.id:elkbee",
  38.                 "time_in_nanos": 163081,
  39.                 "breakdown": {
  40.                   "set_min_competitive_score_count": 0,
  41.                   "match_count": 0,
  42.                   "shallow_advance_count": 0,
  43.                   "set_min_competitive_score": 0,
  44.                   "next_doc": 2447,
  45.                   "match": 0,
  46.                   "next_doc_count": 4,
  47.                   "score_count": 4,
  48.                   "compute_max_score_count": 0,
  49.                   "compute_max_score": 0,
  50.                   "advance": 3552,
  51.                   "advance_count": 1,
  52.                   "score": 5027,
  53.                   "build_scorer_count": 2,
  54.                   "create_weight": 107840,
  55.                   "shallow_advance": 0,
  56.                   "create_weight_count": 1,
  57.                   "build_scorer": 44215
  58.                 }
  59.               }
  60.             ],
  61.             "rewrite_time": 4769,
  62.             "collector": [
  63.               {
  64.                 "name": "MultiCollector",
  65.                 "reason": "search_multi",
  66.                 "time_in_nanos": 1945072,
  67.                 "children": [
  68.                   {
  69.                     "name": "FilteredCollector",
  70.                     "reason": "search_post_filter",
  71.                     "time_in_nanos": 500850,
  72.                     "children": [
  73.                       {
  74.                         "name": "SimpleTopScoreDocCollector",
  75.                         "reason": "search_top_hits",
  76.                         "time_in_nanos": 22577
  77.                       }
  78.                     ]
  79.                   },
  80.                   {
  81.                     "name": "MultiBucketCollector: [[my_scoped_agg, my_global_agg]]",
  82.                     "reason": "aggregation",
  83.                     "time_in_nanos": 867617
  84.                   }
  85.                 ]
  86.               }
  87.             ]
  88.           }
  89.         ],
  90.         "aggregations": [...], 
  91.         "fetch": {...}
  92.       }
  93.     ]
  94.   }
  95. }

The “aggregations” portion has been omitted because it will be covered in the next section.

As you can see, the output is significantly more verbose than before. All the major portions of the query are represented:

  1. The first TermQuery (user.id:elkbee) represents the main term query.
  2. The second TermQuery (message:search) represents the post_filter query.

The Collector tree is fairly straightforward, showing how a single CancellableCollector wraps a MultiCollector which also wraps a FilteredCollector to execute the post_filter (and in turn wraps the normal scoring SimpleCollector), a BucketCollector to run all scoped aggregations.

Understanding MultiTermQuery output

A special note needs to be made about the MultiTermQuery class of queries. This includes wildcards, regex, and fuzzy queries. These queries emit very verbose responses, and are not overly structured.

Essentially, these queries rewrite themselves on a per-segment basis. If you imagine the wildcard query b*, it technically can match any token that begins with the letter “b”. It would be impossible to enumerate all possible combinations, so Lucene rewrites the query in context of the segment being evaluated, e.g., one segment may contain the tokens [bar, baz], so the query rewrites to a BooleanQuery combination of “bar” and “baz”. Another segment may only have the token [bakery], so the query rewrites to a single TermQuery for “bakery”.

Due to this dynamic, per-segment rewriting, the clean tree structure becomes distorted and no longer follows a clean “lineage” showing how one query rewrites into the next. At present time, all we can do is apologize, and suggest you collapse the details for that query’s children if it is too confusing. Luckily, all the timing statistics are correct, just not the physical layout in the response, so it is sufficient to just analyze the top-level MultiTermQuery and ignore its children if you find the details too tricky to interpret.

Hopefully this will be fixed in future iterations, but it is a tricky problem to solve and still in-progress. :)

Profiling Aggregations

aggregations Section

The aggregations section contains detailed timing of the aggregation tree executed by a particular shard. The overall structure of this aggregation tree will resemble your original Elasticsearch request. Let’s execute the previous query again and look at the aggregation profile this time:

  1. GET /my-index-000001/_search
  2. {
  3.   "profile": true,
  4.   "query": {
  5.     "term": {
  6.       "user.id": {
  7.         "value": "elkbee"
  8.       }
  9.     }
  10.   },
  11.   "aggs": {
  12.     "my_scoped_agg": {
  13.       "terms": {
  14.         "field": "http.response.status_code"
  15.       }
  16.     },
  17.     "my_global_agg": {
  18.       "global": {},
  19.       "aggs": {
  20.         "my_level_agg": {
  21.           "terms": {
  22.             "field": "http.response.status_code"
  23.           }
  24.         }
  25.       }
  26.     }
  27.   },
  28.   "post_filter": {
  29.     "match": {
  30.       "message": "search"
  31.     }
  32.   }
  33. }

This yields the following aggregation profile output:

  1. {
  2.   "profile": {
  3.     "shards": [
  4.       {
  5.         "aggregations": [
  6.           {
  7.             "type": "NumericTermsAggregator",
  8.             "description": "my_scoped_agg",
  9.             "time_in_nanos": 79294,
  10.             "breakdown": {
  11.               "reduce": 0,
  12.               "build_aggregation": 30885,
  13.               "build_aggregation_count": 1,
  14.               "initialize": 2623,
  15.               "initialize_count": 1,
  16.               "reduce_count": 0,
  17.               "collect": 45786,
  18.               "collect_count": 4,
  19.               "build_leaf_collector": 18211,
  20.               "build_leaf_collector_count": 1,
  21.               "post_collection": 929,
  22.               "post_collection_count": 1
  23.             },
  24.             "debug": {
  25.               "total_buckets": 1,
  26.               "result_strategy": "long_terms",
  27.               "built_buckets": 1
  28.             }
  29.           },
  30.           {
  31.             "type": "GlobalAggregator",
  32.             "description": "my_global_agg",
  33.             "time_in_nanos": 104325,
  34.             "breakdown": {
  35.               "reduce": 0,
  36.               "build_aggregation": 22470,
  37.               "build_aggregation_count": 1,
  38.               "initialize": 12454,
  39.               "initialize_count": 1,
  40.               "reduce_count": 0,
  41.               "collect": 69401,
  42.               "collect_count": 4,
  43.               "build_leaf_collector": 8150,
  44.               "build_leaf_collector_count": 1,
  45.               "post_collection": 1584,
  46.               "post_collection_count": 1
  47.             },
  48.             "debug": {
  49.               "built_buckets": 1
  50.             },
  51.             "children": [
  52.               {
  53.                 "type": "NumericTermsAggregator",
  54.                 "description": "my_level_agg",
  55.                 "time_in_nanos": 76876,
  56.                 "breakdown": {
  57.                   "reduce": 0,
  58.                   "build_aggregation": 13824,
  59.                   "build_aggregation_count": 1,
  60.                   "initialize": 1441,
  61.                   "initialize_count": 1,
  62.                   "reduce_count": 0,
  63.                   "collect": 61611,
  64.                   "collect_count": 4,
  65.                   "build_leaf_collector": 5564,
  66.                   "build_leaf_collector_count": 1,
  67.                   "post_collection": 471,
  68.                   "post_collection_count": 1
  69.                 },
  70.                 "debug": {
  71.                   "total_buckets": 1,
  72.                   "result_strategy": "long_terms",
  73.                   "built_buckets": 1
  74.                 }
  75.               }
  76.             ]
  77.           }
  78.         ]
  79.       }
  80.     ]
  81.   }
  82. }

From the profile structure we can see that the my_scoped_agg is internally being run as a NumericTermsAggregator (because the field it is aggregating, http.response.status_code, is a numeric field). At the same level, we see a GlobalAggregator which comes from my_global_agg. That aggregation then has a child NumericTermsAggregator which comes from the second term’s aggregation on http.response.status_code.

The time_in_nanos field shows the time executed by each aggregation, and is inclusive of all children. While the overall time is useful, the breakdown field will give detailed stats about how the time was spent.

Some aggregations may return expert debug information that describe features of the underlying execution of the aggregation that are ‘useful for folks that hack on aggregations but that we don’t expect to be otherwise useful. They can vary wildly between versions, aggregations, and aggregation execution strategies.

Timing Breakdown

The breakdown component lists detailed statistics about low-level execution:

  1. "breakdown": {
  2.   "reduce": 0,
  3.   "build_aggregation": 30885,
  4.   "build_aggregation_count": 1,
  5.   "initialize": 2623,
  6.   "initialize_count": 1,
  7.   "reduce_count": 0,
  8.   "collect": 45786,
  9.   "collect_count": 4,
  10.   "build_leaf_collector": 18211,
  11.   "build_leaf_collector_count": 1
  12. }

Each property in the breakdown component corresponds to an internal method for the aggregation. For example, the build_leaf_collector property measures nanoseconds spent running the aggregation’s getLeafCollector() method. Properties ending in _count record the number of invocations of the particular method. For example, "collect_count": 2 means the aggregation called the collect() on two different documents. The reduce property is reserved for future use and always returns 0.

Timings are listed in wall-clock nanoseconds and are not normalized at all. All caveats about the overall time apply here. The intention of the breakdown is to give you a feel for A) what machinery in Elasticsearch is actually eating time, and B) the magnitude of differences in times between the various components. Like the overall time, the breakdown is inclusive of all children times.

Profiling Fetch

All shards the fetched documents will have a fetch section in the profile. Let’s execute a small search and have a look a the fetch profile:

  1. GET /my-index-000001/_search?filter_path=profile.shards.fetch
  2. {
  3.   "profile": true,
  4.   "query": {
  5.     "term": {
  6.       "user.id": {
  7.         "value": "elkbee"
  8.       }
  9.     }
  10.   }
  11. }

And here is the fetch profile:

  1. {
  2.   "profile": {
  3.     "shards": [
  4.       {
  5.         "fetch": {
  6.           "type": "fetch",
  7.           "description": "",
  8.           "time_in_nanos": 660555,
  9.           "breakdown": {
  10.             "next_reader": 7292,
  11.             "next_reader_count": 1,
  12.             "load_stored_fields": 299325,
  13.             "load_stored_fields_count": 5
  14.           },
  15.           "debug": {
  16.             "stored_fields": ["_id", "_routing", "_source"]
  17.           },
  18.           "children": [
  19.             {
  20.               "type": "FetchSourcePhase",
  21.               "description": "",
  22.               "time_in_nanos": 20443,
  23.               "breakdown": {
  24.                 "next_reader": 745,
  25.                 "next_reader_count": 1,
  26.                 "process": 19698,
  27.                 "process_count": 5
  28.               },
  29.               "debug": {
  30.                 "fast_path": 4
  31.               }
  32.             }
  33.           ]
  34.         }
  35.       }
  36.     ]
  37.   }
  38. }

Since this is debugging information about the way that Elasticsearch executes the fetch it can change from request to request and version to version. Even patch versions may change the output here. That lack of consistency is what makes it useful for debugging.

Anyway! time_in_nanos measures the time total time of the fetch phase. The breakdown counts and times the our per-segment preparation in next_reader and the time taken loading stored fields in load_stored_fields. Debug contains miscellaneous non-timing information, specifically stored_fields lists the stored fields that fetch will have to load. If it is an empty list then fetch will entirely skip loading stored fields.

The children section lists the sub-phases that do the actual fetching work and the breakdown has counts and timings for the per-segment preparation in next_reader and the per document fetching in process.

We try hard to load all of the stored fields that we will need for the fetch up front. This tends to make the _source phase a couple of microseconds per hit. In that case the true cost of _source phase is hidden in the load_stored_fields component of the breakdown. It’s possible to entirely skip loading stored fields by setting "_source": false, "stored_fields": ["_none_"].

Profiling Considerations

Like any profiler, the Profile API introduces a non-negligible overhead to search execution. The act of instrumenting low-level method calls such as collect, advance, and next_doc can be fairly expensive, since these methods are called in tight loops. Therefore, profiling should not be enabled in production settings by default, and should not be compared against non-profiled query times. Profiling is just a diagnostic tool.

There are also cases where special Lucene optimizations are disabled, since they are not amenable to profiling. This could cause some queries to report larger relative times than their non-profiled counterparts, but in general should not have a drastic effect compared to other components in the profiled query.

Limitations

  • Profiling currently does not measure the network overhead.
  • Profiling also does not account for time spent in the queue, merging shard responses on the coordinating node, or additional work such as building global ordinals (an internal data structure used to speed up search).
  • Profiling statistics are currently not available for suggestions, highlighting, dfs_query_then_fetch.
  • Profiling of the reduce phase of aggregation is currently not available.

  • The Profiler is instrumenting internals that can change from version to version. The resulting json should be considered mostly unstable, especially things in the debug section.

  • « Explain API

  • Field capabilities API »