[
https://issues.apache.org/jira/browse/FLINK-31238?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17698227#comment-17698227
]
Yue Ma commented on FLINK-31238:
--------------------------------
[~yunta] Thanks for replying ~
Yes, in the current implementation both Clip and Ingest happen in
#initialization phase . But compared to the previous rescaling recovery method,
its blocking time will be much reduced.
I also thought about asynchronous recovery during rescaling. For example,
delete data that does not belong to the Task KeyGroup through a Special
Compaction, and ensure that the compaction can be executed at the first
checkpoint. Or make some special marks for invalid data like DeteleRange and
then clean it up asynchronously.
Considering that most of our online jobs use rocksdb statebackend , we need to
support these feature on rocksdb. It seems that some of other solutions will
bring more changes to rocksdb. Both consider the risks and benefits, we chose
the above method
> Use IngestDB to speed up Rocksdb rescaling recovery
> ----------------------------------------------------
>
> Key: FLINK-31238
> URL: https://issues.apache.org/jira/browse/FLINK-31238
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / State Backends
> Affects Versions: 1.16.1
> Reporter: Yue Ma
> Priority: Major
> Attachments: image-2023-02-27-16-41-18-552.png,
> image-2023-02-27-16-57-18-435.png, image-2023-03-07-14-27-10-260.png,
> image-2023-03-09-15-23-30-581.png, image-2023-03-09-15-26-12-314.png,
> image-2023-03-09-15-28-32-363.png, image-2023-03-09-15-41-03-074.png,
> image-2023-03-09-15-41-08-379.png, image-2023-03-09-15-45-56-081.png,
> image-2023-03-09-15-46-01-176.png, image-2023-03-09-15-50-04-281.png
>
>
> There have been many discussions and optimizations in the community about
> optimizing rocksdb scaling and recovery.
> https://issues.apache.org/jira/browse/FLINK-17971
> https://issues.apache.org/jira/browse/FLINK-8845
> https://issues.apache.org/jira/browse/FLINK-21321
> We hope to discuss some of our explorations under this ticket
> The process of scaling and recovering in rocksdb simply requires two steps
> # Insert the valid keyGroup data of the new task.
> # Delete the invalid data in the old stateHandle.
> The current method for data writing is to specify the main Db first and then
> insert data using writeBatch.In addition, the method of deleteRange is
> currently used to speed up the ClipDB. But in our production environment, we
> found that the speed of rescaling is still very slow, especially when the
> state of a single Task is large.
>
> We hope that the previous sst file can be reused directly when restoring
> state, instead of retraversing the data. So we made some attempts to optimize
> it in our internal version of flink and frocksdb.
>
> We added two APIs *ClipDb* and *IngestDb* in frocksdb.
> * ClipDB is used to clip the data of a DB. Different from db.DeteleRange and
> db.Delete, DeleteValue and RangeTombstone will not be generated for parts
> beyond the key range. We will iterate over the FileMetaData of db. Process
> each sst file. There are three situations here.
> If all the keys of a file are required, we will keep the sst file and do
> nothing
> If all the keys of the sst file exceed the specified range, we will delete
> the file directly.
> If we only need some part of the sst file, we will rewrite the required keys
> to generate a new sst file。
> All sst file changes will be placed in a VersionEdit, and the current
> versions will LogAndApply this edit to ensure that these changes can take
> effect
> * IngestDb is used to directly ingest all sst files of one DB into another
> DB. But it is necessary to strictly ensure that the keys of the two DBs do
> not overlap, which is easy to do in the Flink scenario. The hard link method
> will be used in the process of ingesting files, so it will be very fast. At
> the same time, the file number of the main DB will be incremented
> sequentially, and the SequenceNumber of the main DB will be updated to the
> larger SequenceNumber of the two DBs.
> When IngestDb and ClipDb are supported, the state restoration logic is as
> follows
> * Open the first StateHandle as the main DB and pause the compaction.
> * Clip the main DB according to the KeyGroup range of the Task with ClipDB
> * Open other StateHandles in sequence as Tmp DB, and perform ClipDb
> according to the KeyGroup range
> * Ingest all tmpDb into the main Db after tmpDb cliped
> * Open the Compaction process of the main DB
> !image-2023-02-27-16-57-18-435.png|width=434,height=152!
> We have done some benchmark tests on the internal Flink version, and the test
> results show that compared with the writeBatch method, the expansion and
> recovery speed of IngestDb can be increased by 5 to 10 times as follows
> (SstFileWriter means uses the recovery method of generating sst files through
> SstFileWriter in parallel）
> * parallelism changes from 4 to 2
> |*TaskStateSize*|*Write_Batch*|*SST_File_Writer*|*Ingest_DB*|
> |500M|Iteration 1: 8.018 s/op
> Iteration 2: 9.551 s/op
> Iteration 3: 7.486 s/op|Iteration 1: 6.041 s/op
> Iteration 2: 5.934 s/op
> Iteration 3: 6.707 s/o|{color:#ff0000}Iteration 1: 3.922 s/op{color}
> {color:#ff0000}Iteration 2: 3.208 s/op{color}
> {color:#ff0000}Iteration 3: 3.096 s/op{color}|
> |1G|Iteration 1: 19.686 s/op
> Iteration 2: 19.402 s/op
> Iteration 3: 21.146 s/op|Iteration 1: 17.538 s/op
> Iteration 2: 16.933 s/op
> Iteration 3: 15.486 s/op|{color:#ff0000}Iteration 1: 6.207 s/op{color}
> {color:#ff0000}Iteration 2: 7.164 s/op{color}
> {color:#ff0000}Iteration 3: 6.397 s/op{color}|
> |5G|Iteration 1: 244.795 s/op
> Iteration 2: 243.141 s/op
> Iteration 3: 253.542 s/op|Iteration 1: 78.058 s/op
> Iteration 2: 85.635 s/op
> Iteration 3: 76.568 s/op|{color:#ff0000}Iteration 1: 23.397 s/op{color}
> {color:#ff0000}Iteration 2: 21.387 s/op{color}
> {color:#ff0000}Iteration 3: 22.858 s/op{color}|
> * parallelism changes from 4 to 8
> |*TaskStateSize*|*Write_Batch*|*SST_File_Writer*|*Ingest_DB*|
> |500M|Iteration 1: 3.477 s/op
> Iteration 2: 3.515 s/op
> Iteration 3: 3.433 s/op|Iteration 1: 3.453 s/op
> Iteration 2: 3.300 s/op
> Iteration 3: 3.313 s/op|{color:#ff0000}Iteration 1: 0.941 s/op{color}
> {color:#ff0000}Iteration 2: 0.963 s/op{color}
> {color:#ff0000}Iteration 3: 1.102 s/op{color}|
> |1G|IIteration 1: 7.571 s/op
> Iteration 2: 7.352 s/op
> Iteration 3: 7.568 s/op|Iteration 1: 5.032 s/op
> Iteration 2: 4.689 s/op
> Iteration 3: 6.883 s/op|{color:#ff0000}Iteration 1: 2.130 s/op{color}
> {color:#ff0000}Iteration 2: 2.110 s/op{color}
> {color:#ff0000}Iteration 3: 2.034 s/op{color}|
> |5G|Iteration 1: 91.870 s/op
> Iteration 2: 94.229 s/op
> Iteration 3: 93.271 s/op|Iteration 1: 25.845 s/op
> Iteration 2: 25.571 s/op
> Iteration 3: 25.685 s/op|{color:#ff0000}Iteration 1: 11.154 s/op{color}
> {color:#ff0000}Iteration 2: 10.732 s/op{color}
> {color:#ff0000}Iteration 3: 10.622 s/op{color}|
> * parallelism changes from 4 to 6
> |*TaskStateSize*|*Write_Batch*|*SST_File_Writer*|*Ingest_DB*|
> |500M|Iteration 1: 8.209 s/op
> Iteration 2: 9.893 s/op
> Iteration 3: 9.150 s/op|Iteration 1: 6.041 s/op
> Iteration 2: 5.934 s/op
> Iteration 3: 6.707 s/o|{color:#ff0000}Iteration 1: 2.622 s/op{color}
> {color:#ff0000}Iteration 2: 2.545 s/op{color}
> {color:#ff0000}Iteration 3: 2.573 s/op{color}|
> |1G|Iteration 1: 21.206 s/op
> Iteration 2: 26.214 s/op
> Iteration 3: 20.269 s/op|Iteration 1: 10.043 s/op
> Iteration 2: 10.744 s/op
> Iteration 3: 10.461 s/op|{color:#ff0000}Iteration 1: 4.400 s/op{color}
> {color:#ff0000}Iteration 2: 4.340 s/op{color}
> {color:#ff0000}Iteration 3: 6.234 s/op{color}|
> |5G|IIteration 1: 170.606 s/op
> Iteration 2: 160.576 s/op
> Iteration 3: 159.425 s/op|IIteration 1: 52.537 s/op
> Iteration 2: 50.576 s/op
> Iteration 3: 50.823 s/op|{color:#ff0000}Iteration 1: 19.053 s/op{color}
> {color:#ff0000}Iteration 2: 18.504 s/op{color}
> {color:#ff0000}Iteration 3: 18.249 s/op{color}|
> * parallelism changes from 4 to 3
> |*TaskStateSize*|*Write_Batch*|*SST_File_Writer*|*Ingest_DB*|
> |500M|Iteration 1: 6.330 s/op
> Iteration 2: 5.614 s/op
> Iteration 3: 5.736 s/op|Iteration 1: 4.083 s/op
> Iteration 2: 5.655 s/op
> Iteration 3: 3.998 s/op|{color:#ff0000}Iteration 1: 2.157 s/op{color}
> {color:#ff0000}Iteration 2: 2.201 s/op{color}
> {color:#ff0000}Iteration 3: 3.212 s/op{color}|
> |1G|Iteration 1: 13.814 s/op
> Iteration 2: 12.852 s/op
> Iteration 3: 13.480 s/op|Iteration 1: 9.619 s/op
> Iteration 2: 9.197 s/op
> Iteration 3: 8.694 s/op|{color:#ff0000}Iteration 1: 4.227 s/op{color}
> {color:#ff0000}Iteration 2: 4.234 s/op{color}
> {color:#ff0000}Iteration 3: 4.177 s/op{color}|
> |5G|Iteration 1: 136.621 s/op
> Iteration 2: 127.097 s/op
> Iteration 3: 139.694 s/op|Iteration 1: 39.612 s/op
> Iteration 2: 38.809 s/op
> Iteration 3: 39.125 s/op|{color:#ff0000}Iteration 1: 16.691 s/op{color}
> {color:#ff0000}Iteration 2: 16.599 s/op{color}
> {color:#ff0000}Iteration 3: 16.726 s/op{color}|
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
