query slow about near real time data

[cxlRay]

when send batch query to Realtime processes tasks, the performance is too bad. TPS only have 80, response time is more than one second,
max response time can be 30 second and 99% response time can be 15 second.
what I confused is the data of Realtime processes are in mem, why response time of query is so long?

Affected Version

v druid-0.16.1-incubating.

Description

• Cluster size
  coordinator and overlord: 2
  historical: 7
  middleManager: 7
  broker: 5

• The testing tool is jmeter

• testing result
  thread num: 500
  TPS: 107.80
  average response time: 4372ms
  99% response time: 15559ms
  max response time: 30150ms
  min response time: 373

• the Flame chart of Realtime task code when test
  

• the configuration of middleManager
  

• the configuration of Realtime tasks

{
  "type": "kafka",
  "dataSchema": {
    "dataSource": "xxxx",
    "parser": {
      "type": "string",
      "parseSpec": {
        "format": "json",
        "timestampSpec": {
          "column": "timestamp",
          "format": "posix"
        },
        "dimensionsSpec": {
          "dimensions": ["tag1","tag2","tag3","tag4","tag5","tag6","tag7"],
          "dimensionExclusions": [
            "timestamp",
            "value"
          ]
        }
      }
    },
    "metricsSpec": [
      {
        "name": "value",
        "fieldName": "value",
        "type": "doubleSum"
      }
    ],
    "granularitySpec": {
      "type": "uniform",
      "segmentGranularity": "HOUR",
      "queryGranularity": "NONE",
      "rollup" : false
    }
  },
  "tuningConfig": {
    "type": "kafka",
    "intermediatePersistPeriod": "PT1H",
    "maxTotalRows": "245000000",
    "maxRowsPerSegment": 5000000
  },
  "ioConfig": {
    "topic": "xxxxx",
    "consumerProperties": {
      "bootstrap.servers": "xxxx:9092"
    },
    "taskCount": 16,
    "replicas": 1,
    "taskDuration": "PT1H"
  }
}

• about segments
  
  

• my query

{
  "queryType": "timeseries",
  "dataSource": "xxxx",
  "granularity": "second",
  "context": {
  	"skipEmptyBuckets": true,
  	"vectorize": "true"
  },
  "filter": { "type": "and", "fields": [{ "type": "selector", "dimension": "endpoint", "value": "host"}, { "type": "selector", "dimension": "metric", "value":"cpu.busy"}] },
  "aggregations": [
  	{ "type": "count", "name": "count"},
  	{ "type": "stringLast", "name": "dsType", "fieldName": "counterType" },
  	{ "type": "doubleMax", "name": "max_value", "fieldName": "value" },
  	{ "type": "doubleMin", "name": "min_value", "fieldName": "value" },
    { "type": "doubleSum", "name": "sum_value", "fieldName": "value" }
  ],
  "postAggregations": [
    { "type": "arithmetic",
      "name": "avg",
      "fn": "/",
      "fields": [
        { "type": "fieldAccess", "name": "sum_value", "fieldName": "sum_value" },
        { "type": "fieldAccess", "name": "count", "fieldName": "count" }
      ]
    }
  ],
  "intervals": [ "2020-05-25T15:35:00+08:00/2020-05-25T15:52:00+08:00" ]
}

[yuanlihan]

Hi @cxlRay,
You can try to minimise intermediatePersistPeriod which is PT10M by default and then the persisted data could apply the "vectorize": "true" option. Hope this helps to some extent.

[exherb]

some issue here.

[cxlRay]

@yuanlihan thanks for your answer, minimise intermediatePersistPeriod will create much more small file

[yuanlihan]

@yuanlihan thanks for your answer, minimise intermediatePersistPeriod will create much more small file

@cxlRay that's true. But the temporary persist files will be cleaned up when the hourly tasks finished. And the persisted incremental files/indexes(with extra indexes to speed up query processing) will be more efficient than the in-memory incrementalIndex. As far as I know, when a query scan the latest in-memory incrementalIndex(like latest 10 minutes's data), Druid processes the in-memory fact table held by incrementalIndex row by row.
Actually, I had tried to minimise the value of maxRowsInMemory to reduce in-memory rows but this will also introduce some overheads caused by frequently persisting.

[exherb]

timeseries is very slow with filter. (10-20s)

[cxlRay]

@exherb say more about your question, if your cluster run on ssd, there will be better

[exherb]

@exherb say more about your question, if your cluster run on ssd, there will be better

• historacal & middlemanager runs on ssd.
• 5million rows per segement / kafka ingestion task
• use turnilo to query
• historical query/time is acceptable (around 100-500ms)
• peon query：timeseries(lookup) / topN query/time is acceptable
• peon query：topN(lookup) / timeseries is very slow (timeseries query/time: 16s-40s)

[cxlRay]

just like yuanlihan say，minimise intermediatePersistPeriod or maxRowsInMemory , add query parmeter vectorize

[exherb]

just like yuanlihan say，minimise intermediatePersistPeriod or maxRowsInMemory , add query parmeter vectorize

intermediatePersistPeriod: P10M
maxRowsInMemory: 1000000

context: { vectorize: true }

still slow.

[navis]

rows in no-rollup incremental index are stored as

ConcurrentSkipListMap<Long, ConcurrentLinkedDeque<IncrementalIndexRow>>

which seemed not easy to be fast, imho.

Anyway, can you try with coarser granularity, like 15 minutes or something?

[exherb]

rows in no-rollup incremental index are stored as

ConcurrentSkipListMap<Long, ConcurrentLinkedDeque<IncrementalIndexRow>>

which seemed not easy to be fast, imho.

Anyway, can you try with coarser granularity, like 15 minutes or something?

we enabled rollup to 1minutes, segment granularity by hour. Are you suggestion change segment granularity to 15 minutes?

[cxlRay]

@exherb how do you know the slow part is peon query,the query process: client -> broker -> indexing service(Peon), use monitor metric or analysis code?

[exherb]

@exherb how do you know the slow part is peon query,the query process: client -> broker -> indexing service(Peon), use monitor metric or analysis code?

by druid metrics: query/segment/time / query/time (datasource~=.+middlemanager.+)

[navis]

@exherb I mean granularity of timeseries query.

[louisliu318]

same issue.