MongoDB – Slow text search query with multiple words on a large dataset - mongodb

Question

I have a M10 Atlas instance.
The issue happens in a DB with 2.5m of documents, but only 900k of { status: "active" } ones.
The search is quick when no sorting/projection is performed, but in order to show the most accurate results first, it needs to be done otherwise the results are a mess.
For some reason this text search is being very slow for medium/long phrases and I have no idea how to speed it up.
Having the following document:
{
_id: "FOO123",
track_date: { $date: "2019-03-09T05:49:22.000Z" },
category_id: "foo",
parent_category_id: "foo",
title: "Disco Ssd Solido Kingston 480g Macbook Pro Air iMac Martinez",
status: "active",
site_id: "foo",
seller_id: 9999,
price: 9999,
permalink:
"https://example.com",
secure_thumbnail:
"https://example.com/img.jpg",
images: [
"https://example.com/img.jpg",
],
fluctuation: 80,
last_update: { $date: "2020-12-11T14:00:24.715Z" },
original_price: null,
}
Performing the following query:
db.articles.aggregate([
{
"$match": {
"status": "active",
"site_id": "foo",
"$text": {
"$search": "Disco Ssd solido kingston 480g macbook pro air",
"$caseSensitive": false
}
}
},
{
"$sort": {
"score": {
"$meta": "textScore"
}
}
},
{
"$skip": 0
},
{
"$limit": 10
}
])
Takes around 35s to yield results.
These are the indexes that I have set up for the active and title fields (I also have compound indexes that mix these two and a few more properties, I can show them here if needed)
{
v: 2,
key: { _fts: "text", _ftsx: 1 },
name: "title_text",
language_override: "language",
weights: { title: 1 },
default_language: "english",
ns: "foo.articles",
textIndexVersion: 3,
},
{
v: 2,
key: { status: 1 },
name: "status_1",
ns: "foo.articles",
}
This is the result of explain("executionStats")
Query (as aggregation did not work with "executionStats":
collection.find(
{
status: "active",
site_id: "foo",
$text: {
$search: "Disco Ssd solido kingston 480g macbook pro air",
$caseSensitive: false,
},
},
{ score: { $meta: "textScore" } }
)
.sort({ score: { $meta: "textScore" } })
.skip(0)
.limit(10)
.explain("executionStats")
Explain with execution stats:
{
queryPlanner: {
plannerVersion: 1,
namespace: "foo.articles",
indexFilterSet: false,
parsedQuery: {
$and: [
{ site_id: { $eq: "foo" } },
{ status: { $eq: "active" } },
{
$text: {
$search: "Disco Ssd solido kingston 480g macbook pro air",
$language: "english",
$caseSensitive: false,
$diacriticSensitive: false,
},
},
],
},
winningPlan: {
stage: "PROJECTION_DEFAULT",
transformBy: { score: { $meta: "textScore" } },
inputStage: {
stage: "SORT",
sortPattern: { score: { $meta: "textScore" } },
limitAmount: 10,
inputStage: {
stage: "SORT_KEY_GENERATOR",
inputStage: {
stage: "FETCH",
filter: {
$and: [
{ site_id: { $eq: "foo" } },
{ status: { $eq: "active" } },
],
},
inputStage: {
stage: "TEXT",
indexPrefix: {},
indexName: "title_text",
parsedTextQuery: {
terms: [
"480g",
"air",
"disco",
"kingston",
"macbook",
"pro",
"solido",
"ssd",
],
negatedTerms: [],
phrases: [],
negatedPhrases: [],
},
textIndexVersion: 3,
inputStage: {
stage: "TEXT_MATCH",
inputStage: {
stage: "TEXT_OR",
inputStages: [
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
{
stage: "IXSCAN",
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
},
],
},
},
},
},
},
},
},
rejectedPlans: [],
},
executionStats: {
executionSuccess: true,
nReturned: 10,
executionTimeMillis: 30458,
totalKeysExamined: 169654,
totalDocsExamined: 279454,
executionStages: {
stage: "PROJECTION_DEFAULT",
nReturned: 10,
executionTimeMillisEstimate: 27307,
works: 309403,
advanced: 10,
needTime: 309392,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
transformBy: { score: { $meta: "textScore" } },
inputStage: {
stage: "SORT",
nReturned: 10,
executionTimeMillisEstimate: 27299,
works: 309403,
advanced: 10,
needTime: 309392,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
sortPattern: { score: { $meta: "textScore" } },
memUsage: 9275,
memLimit: 33554432,
limitAmount: 10,
inputStage: {
stage: "SORT_KEY_GENERATOR",
nReturned: 31606,
executionTimeMillisEstimate: 27262,
works: 309392,
advanced: 31606,
needTime: 277785,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
inputStage: {
stage: "FETCH",
filter: {
$and: [
{ site_id: { $eq: "foo" } },
{ status: { $eq: "active" } },
],
},
nReturned: 31606,
executionTimeMillisEstimate: 27251,
works: 309391,
advanced: 31606,
needTime: 277784,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
docsExamined: 139727,
alreadyHasObj: 139727,
inputStage: {
stage: "TEXT",
nReturned: 139727,
executionTimeMillisEstimate: 27191,
works: 309391,
advanced: 139727,
needTime: 169663,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
indexPrefix: {},
indexName: "title_text",
parsedTextQuery: {
terms: [
"480g",
"air",
"disco",
"kingston",
"macbook",
"pro",
"solido",
"ssd",
],
negatedTerms: [],
phrases: [],
negatedPhrases: [],
},
textIndexVersion: 3,
inputStage: {
stage: "TEXT_MATCH",
nReturned: 139727,
executionTimeMillisEstimate: 27191,
works: 309391,
advanced: 139727,
needTime: 169663,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
docsRejected: 0,
inputStage: {
stage: "TEXT_OR",
nReturned: 139727,
executionTimeMillisEstimate: 27191,
works: 309391,
advanced: 139727,
needTime: 169663,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
docsExamined: 139727,
inputStages: [
{
stage: "IXSCAN",
nReturned: 291,
executionTimeMillisEstimate: 0,
works: 292,
advanced: 291,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 291,
seeks: 1,
dupsTested: 291,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 26973,
executionTimeMillisEstimate: 121,
works: 26974,
advanced: 26973,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 26973,
seeks: 1,
dupsTested: 26973,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 29687,
executionTimeMillisEstimate: 69,
works: 29688,
advanced: 29687,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 29687,
seeks: 1,
dupsTested: 29687,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 13595,
executionTimeMillisEstimate: 28,
works: 13596,
advanced: 13595,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 13595,
seeks: 1,
dupsTested: 13595,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 10730,
executionTimeMillisEstimate: 28,
works: 10731,
advanced: 10730,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 10730,
seeks: 1,
dupsTested: 10730,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 53358,
executionTimeMillisEstimate: 248,
works: 53359,
advanced: 53358,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 53358,
seeks: 1,
dupsTested: 53358,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 7847,
executionTimeMillisEstimate: 13,
works: 7848,
advanced: 7847,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 7847,
seeks: 1,
dupsTested: 7847,
dupsDropped: 0,
},
{
stage: "IXSCAN",
nReturned: 27173,
executionTimeMillisEstimate: 17,
works: 27174,
advanced: 27173,
needTime: 0,
needYield: 0,
saveState: 3211,
restoreState: 3211,
isEOF: 1,
keyPattern: { _fts: "text", _ftsx: 1 },
indexName: "title_text",
isMultiKey: true,
isUnique: false,
isSparse: false,
isPartial: false,
indexVersion: 2,
direction: "backward",
indexBounds: {},
keysExamined: 27173,
seeks: 1,
dupsTested: 27173,
dupsDropped: 0,
},
],
},
},
},
},
},
},
},
},
serverInfo: {
host: "host",
port: "port",
version: "4.2.11",
gitVersion: "ea38428f0c6742c7c2c7f677e73d79e17a2aab96",
},
ok: 1,
$clusterTime: {
clusterTime: Timestamp(1, 1609902178),
signature: {
hash: BinData(0, "cunb+7FEjXmbwN22uz3F+nV0LKE="),
keyId: NumberLong("6911938552179720195"),
},
},
operationTime: Timestamp(1, 1609902178),
}
The issue is not related to the status field, as I've tried removing it and the issue persists.
I don't really mind a query that's a bit slow, but more than half a minute is too much for a search.
At first I was doing an AND query for the text search instead of an OR, but it was even slower for long phrases, so decided to give the score a try.
I've also tried removing the sort pipeline, but it still takes 30s~
I'll take any tip anyone can give me at this point.
EDIT: According to the explain("executionStats") bit, it seems that the sort/projection of score is the time sink, but I've no idea how to improve that in order to get the best matches first

Answer 1

Mongodb Atlas specific feature
Since you are using Atlas you may find https://docs.atlas.mongodb.com/atlas-search a better alternative. It uses Lucene engine similar to Elasticsearch, Solr, etc, instead of rather limited native text index which is available on self-hosted versions of the database.
Atlas search is way more flexible and in most cases faster than mongo text index. The downside is a noticeable lag (like very eventual consistency), and higher demand on disk space (especially for ngram tokenizers) but these are not critical in your case.
The search is available on clusters of v4.2+:
Give it a try, it's included in your subscription.
The highlighted parts of the question must be confusing and are deserved some explanations:
The issue happens in a DB with 2.5m of documents, but only 900k of { status: "active" } ones.
It doesn't really matter how many active documents there are. There is no index intersection so the only text index is used. Filter by status and site_id is done only on the FETCH stage. In other words you search all 2.5m documents regardless of other filters.
The search is quick when no sorting/projection is performed
$sort is a blocking stage, means it requires previous stage to complete. You have $skip-$limit stages to return 10 documents only. Without $sort stage the pipeline returns 10 documents that were found first. With $sort stage the pipeline waits for all documents to be searched to find 10 with highest score.
For some reason this text search is being very slow for medium/long phrases
The reason is obvious - the more words in the phrase the more searches mongo need to do. It's one search per word, then calculation of the score based on the results of all searches.
Memory shouldn't be an issue for the sorting stage. The limit is 100 MB, then it starts using disk regardless of how much RAM you have.
Lack of memory might be a problem if all indexes and the working set don't fit into RAM. If it is the case it should be highlighted in the "Performance Advisor" tab.
db.collection.totalIndexSize() can also give you some idea of how much RAM you would need for optimal performance. It is not specific to the text index and affects performance of the whole cluster.