[FLINK-26050] Too many small sst files in rocksdb state backend when using time window created in ascending order

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.10.2, 1.14.3
• Fix Version/s: 1.20.0
• Component/s: Runtime / State Backends
• Labels: pull-request-available

Release Note

In some cases, the number of files produced by RocksDB state backend grows indefinitely.This might cause task state info (TDD and checkpoint ACK) to exceed RPC message size and fail recovery/checkpoint in addition to having lots of small files.
In Flink 1.20, you can manually merge such files in the background using RocksDB API.

• Assignee: Roman Khachatryan
• Reporter: shen
• Created: 09/Feb/22 10:21
• Updated: 02/Aug/24 14:32
• Resolved: 02/Aug/24 14:32

Description

When using processing or event time windows, in some workloads, there will be a lot of small sst files(serveral KB) in rocksdb local directory and may cause "Too many files error".

Use rocksdb tool ldb to find out content in sst files:

• column family of these small sst files is "processing_window-timers".
• most sst files are in level-1.
• records in sst files are almost kTypeDeletion.
• creation time of sst file correspond to checkpoint interval.

These small sst files seem to be generated when flink checkpoint is triggered. Although all content in sst are delete tags, they are not compacted and deleted in rocksdb compaction because of not intersecting with each other(rocksdb compaction trivial move). And there seems to be no chance to delete them because of small size and not intersect with other sst files.

 

I will attach a simple program to reproduce the problem.

 

Since timer in processing time window is generated in strictly ascending order(both put and delete). So If workload of job happen to generate level-0 sst files not intersect with each other.(for example: processing window size much smaller than checkpoint interval, and no window content cross checkpoint interval or no new data in window crossing checkpoint interval). There will be many small sst files generated until job restored from savepoint, or incremental checkpoint is disabled. 

 

May be similar problem exists when user use timer in operators with same workload.

 

Code to reproduce the problem:

package org.apache.flink.jira;

import lombok.extern.slf4j.Slf4j;
import org.apache.flink.configuration.Configuration;
import org.apache.flink.configuration.RestOptions;
import org.apache.flink.configuration.TaskManagerOptions;
import org.apache.flink.contrib.streaming.state.RocksDBStateBackend;
import org.apache.flink.streaming.api.TimeCharacteristic;
import org.apache.flink.streaming.api.checkpoint.ListCheckpointed;
import org.apache.flink.streaming.api.datastream.DataStreamSource;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.source.SourceFunction;
import org.apache.flink.streaming.api.functions.windowing.ProcessWindowFunction;
import org.apache.flink.streaming.api.windowing.assigners.TumblingProcessingTimeWindows;
import org.apache.flink.streaming.api.windowing.time.Time;
import org.apache.flink.streaming.api.windowing.windows.TimeWindow;
import org.apache.flink.util.Collector;

import java.util.Collections;
import java.util.List;
import java.util.Random;

@Slf4j
public class StreamApp  {
  public static void main(String[] args) throws Exception {
    Configuration config = new Configuration();
    config.set(RestOptions.ADDRESS, "127.0.0.1");
    config.set(RestOptions.PORT, 10086);
    config.set(TaskManagerOptions.NUM_TASK_SLOTS, 6);
    new StreamApp().configureApp(StreamExecutionEnvironment.createLocalEnvironment(1, config));
  }

  public void configureApp(StreamExecutionEnvironment env) throws Exception {

    env.enableCheckpointing(20000); // 20sec

    RocksDBStateBackend rocksDBStateBackend =
        new RocksDBStateBackend("file:///Users/shenjiaqi/Workspace/jira/flink-51/checkpoints/", true); // need to be reconfigured
    rocksDBStateBackend.setDbStoragePath("/Users/shenjiaqi/Workspace/jira/flink-51/flink/rocksdb_local_db"); // need to be reconfigured

    env.setStateBackend(rocksDBStateBackend);
    env.getCheckpointConfig().setCheckpointTimeout(100000);
    env.getCheckpointConfig().setTolerableCheckpointFailureNumber(5);
    env.setParallelism(1);
    env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);

    env.getConfig().setTaskCancellationInterval(10000);

    for (int i = 0; i < 1; ++i) {
      createOnePipeline(env);
    }

    env.execute("StreamApp");
  }


  private void createOnePipeline(StreamExecutionEnvironment env) {
    // data source is configured so that little window cross checkpoint interval
    DataStreamSource<String> stream = env.addSource(new Generator(1, 3000, 3600));

    stream.keyBy(x -> x)
        // make sure window size less than checkpoint interval. though 100ms is too small, I think increase this value can still reproduce the problem with longer time.
        .window(TumblingProcessingTimeWindows.of(Time.milliseconds(100)))
        .process(new ProcessWindowFunction<String, String, String, TimeWindow>() {
          @Override
          public void process(String s, ProcessWindowFunction<String, String, String, TimeWindow>.Context context,
              Iterable<String> elements, Collector<String> out) {
            for (String ele: elements) {
              out.collect(ele);
            }
          }
        }).print();
  }

  public static final class Generator implements SourceFunction<String>, ListCheckpointed<Integer> {

    private static final long serialVersionUID = -2819385275681175792L;

    private final int numKeys;
    private final int idlenessMs;
    private final int recordsToEmit;

    private volatile int numRecordsEmitted = 0;
    private volatile boolean canceled = false;

    Generator(final int numKeys, final int idlenessMs, final int durationSeconds) {
      this.numKeys = numKeys;
      this.idlenessMs = idlenessMs;

      this.recordsToEmit = ((durationSeconds * 1000) / idlenessMs) * numKeys;
    }

    @Override
    public void run(final SourceContext<String> ctx) throws Exception {
      Random rnd = new Random();
      while (numRecordsEmitted < recordsToEmit) {
        synchronized (ctx.getCheckpointLock()) {
          for (int i = 0; i < numKeys; i++) {
            ctx.collect("" + rnd.nextInt(1));
            numRecordsEmitted++;
          }
        }
        Thread.sleep(idlenessMs);
      }

      while (!canceled) {
        Thread.sleep(50);
      }

    }

    @Override
    public void cancel() {
      canceled = true;
    }

    @Override
    public List<Integer> snapshotState(final long checkpointId, final long timestamp) {
      return Collections.singletonList(numRecordsEmitted);
    }

    @Override
    public void restoreState(final List<Integer> states) {
      for (final Integer state : states) {
        numRecordsEmitted += state;
      }
    }
  }
}
 

 

 Code to simulate flink checkpointing and timer creation and deletion and reproduce the problem:

//
//  main.cpp
//  reproduce
//
//  Created by shenjiaqi on 2022/2/8.
//

#include <iostream>
#include <filesystem>
#include <cstdio>
#include <cstdlib>
#include <string>

#include "rocksdb/utilities/checkpoint.h"
#include "rocksdb/db.h"
#include "rocksdb/slice.h"
#include "rocksdb/options.h"

using namespace ROCKSDB_NAMESPACE;
using ROCKSDB_NAMESPACE::DB;
using ROCKSDB_NAMESPACE::Options;
using ROCKSDB_NAMESPACE::PinnableSlice;
using ROCKSDB_NAMESPACE::ReadOptions;
using ROCKSDB_NAMESPACE::Status;
using ROCKSDB_NAMESPACE::WriteBatch;
using ROCKSDB_NAMESPACE::WriteOptions;

std::string kDBPath = "/Users/shenjiaqi/Workspace/flink/jira/data-test"; // need to be reconfigured

static void createCheckpoint(rocksdb::DB *db, rocksdb::Status &s) {
    std::cout << "create checkpoint" << std::endl;
    std::string chkPath = kDBPath + "-chp";
    assert(chkPath.find("/Users/shenjiaqi/Workspace/flink/jira/") >= 0); // just in case
    system(("rm -rf " + chkPath).data()); // use with care.
    
    Checkpoint* checkpoint_ptr;
    s = Checkpoint::Create(db, &checkpoint_ptr);
    assert(s.ok());
    
    s = checkpoint_ptr->CreateCheckpoint(chkPath);
    assert(s.ok());
}


int main() {
    DB* db;
    Options options;
    // Optimize RocksDB. This is the easiest way to get RocksDB to perform well
    options.IncreaseParallelism();
    options.OptimizeLevelStyleCompaction();
    // create the DB if it's not already present
    options.create_if_missing = true;
    options.info_log_level = DEBUG_LEVEL;
//    options.level_compaction_dynamic_level_bytes = true;

    // open DB
    Status s = DB::Open(options, kDBPath, &db);
    assert(s.ok());

    for (int i = 0; i < 1000; ++i) {

        std::string key = "key" + /* std::to_string((int)rand()); // */std::to_string(i);
        std::string value = "value" + std::to_string(i);

        // Put key-value
        s = db->Put(WriteOptions(), key, value);
        assert(s.ok());

        // delete after put
        s = db->Delete(WriteOptions(), key);
        assert(s.ok());

        if (i > 0 && (i % 5) == 0) {
            createCheckpoint(db, s);
        }
    }
    
    createCheckpoint(db, s);
    return 0;
}

Many log such as: "Moving #407 to level-1 1047 bytes" can be found in LOG of rocksdb (not enabled in flink by default).

Issue Links

• is related to: FLINK-26800 write small data file using share write buffer manager
• relates to: FLINK-34430 Akka frame size exceeded with many ByteStreamStateHandle being used
• relates to: FLINK-35970 Update documentation about FLINK-26050 (merging small SST files)
• links to: GitHub Pull Request #24880