Fabian Hueske commented on FLINK-8601:

Thanks for the detailed design document [~sihuazhou]! 
Now I have a pretty good understanding of the proposal. 

I think this would be a nice feature, but I'm not sure if it is generic enough 
to go into the base classes that are used by all functions or whether it would 
make more sense to expose it only to specific functions like the 

[~aljoscha], what do you think about the proposal?

> Introduce PartitionedBloomFilter for Approximate calculation and other 
> situations of performance optimization
> -------------------------------------------------------------------------------------------------------------
>                 Key: FLINK-8601
>                 URL: https://issues.apache.org/jira/browse/FLINK-8601
>             Project: Flink
>          Issue Type: New Feature
>          Components: DataStream API, State Backends, Checkpointing
>    Affects Versions: 1.4.0
>            Reporter: Sihua Zhou
>            Assignee: Sihua Zhou
>            Priority: Major
> h3. Backgroud
> Bloom filter is useful in many situation, for example:
>  * 1. Approximate calculation: deduplication (eg: UV calculation)
>  * 2. Performance optimization: eg, [runtime filter 
> join|https://www.cloudera.com/documentation/enterprise/5-9-x/topics/impala_runtime_filtering.html]
>        By using BF, we can greatly reduce the number of queries for state 
> data in a stream join, and these filtered queries will eventually fail to 
> find any results, which is a poor performance for rocksdb-based state due to 
> traversing ```sst``` on the disk. 
> However, based on the current status provided by flink, it is hard to use the 
> bloom filter for the following reasons:
>  * 1. Serialization problem: Bloom filter status can be large (for example: 
> 100M), if implement it based on the RocksDB state, the state data will need 
> to be serialized each time it is queried and updated, and the performance 
> will be very poor.
>  * 2. Data skewed: Data in different key group can be skewed, and the 
> information of data skewed can not be accurately predicted before the program 
> is running. Therefore, it is impossible to determine how much resources bloom 
> filter should allocate. One way to do this is to allocate space needed for 
> the most skewed case, but this can lead to very serious waste of resources.
> h3. Requirement
> Therefore, I introduce the PartitionedBloomFilter for flink, which at least 
> need to meet the following features:
>  * 1. Support for changing Parallelism
>  * 2. Only serialize when necessary: when performing checkpoint
>  * 3. Can deal with data skew problem: users only need to specify a 
> PartitionedBloomFilter with the desired input, fpp, system will allocate 
> resource dynamic.
>  * 4. Do not conflict with other state: user can use KeyedState and 
> OperateState when using this bloom filter.
>  * 5. Support relax ttl (ie: the data survival time at least greater than the 
> specified time)
> Design doc:  [design 
> doc|https://docs.google.com/document/d/1s8w2dkNFDM9Fb2zoHwHY0hJRrqatAFta42T97nDXmqc/edit?usp=sharing]

This message was sent by Atlassian JIRA

Reply via email to