Sihua Zhou commented on FLINK-9070:

Hi [~kien_truong] I also share my idea for improving this here

It base on the follow things
- Using RocksIterator to iterate the records directly (currently we use the 
RocksDBMapIterator, witch will buffer the records and may also need to perform 
seeking multi times.)
- Using WriteBatch to perform deleting in bulk.

What do you think?

> Improve performance of RocksDBMapState.clear()
> ----------------------------------------------
>                 Key: FLINK-9070
>                 URL: https://issues.apache.org/jira/browse/FLINK-9070
>             Project: Flink
>          Issue Type: Improvement
>          Components: State Backends, Checkpointing
>    Affects Versions: 1.6.0
>            Reporter: Truong Duc Kien
>            Priority: Minor
> Currently, RocksDBMapState.clear() is implemented by iterating over all the 
> keys and drop them one by one. This iteration can be quite slow with: 
>  * Large maps
>  * High-churn maps with a lot of tombstones
> There are a few methods to speed-up deletion for a range of keys, each with 
> their own caveats:
>  * DeleteRange: still experimental, likely buggy
>  * DeleteFilesInRange + CompactRange: only good for large ranges
> Flink can also keep a list of inserted keys in-memory, then directly delete 
> them without having to iterate over the Rocksdb database again. 
> Reference:
>  * [RocksDB article about range 
> deletion|https://github.com/facebook/rocksdb/wiki/Delete-A-Range-Of-Keys]
>  * [Bug in DeleteRange|https://pingcap.com/blog/2017-09-08-rocksdbbug]

This message was sent by Atlassian JIRA

Reply via email to