Here is my "little harsh/straightforward feedback", but it's based on fact and real-world experience with using Redis since ~2012.
Redis is not a database, period. The best description of what Redis is is something along the lines of "in-memory - text only (base64 ftw) - data structures on top of TCP socket". The sweet spot for Redis is the in-memory caching layer (Memcache is the closest equivalent). The misconceptions around Redis have haunted me for the past decade. Redis is not providing any good primitives for participating in the checkpointing / fault-tolerance mechanism, beyond what can be implemented natively in the heap-state backend. It can do a full snapshot of a database (costly, we need incremental, ...) or a text-based append log (changelog state backend). All data needs to fit in memory. In text form. No compression (you can of course compress before doing base64). We should not even bother comparing it to RocksDB, its Flink equivalent is the HeapStateBackend, which could be made way more performant because it eliminates the need of going through the network stack. Best, D. On Wed, Jan 31, 2024 at 5:27 PM David Anderson <dander...@apache.org> wrote: > When it comes to decoupling the state store from Flink, I suggest taking a > look at FlinkNDB, which is an experimental state backend for Flink that > puts the state into an external distributed database. There's a Flink > Forward talk [1] and a master's thesis [2] available. > > [1] https://www.youtube.com/watch?v=ZWq_TzsXssM > [2] http://www.diva-portal.org/smash/get/diva2:1536373/FULLTEXT01.pdf > > > > > > > On Wed, Jan 31, 2024 at 12:30 AM Chirag Dewan via user < > user@flink.apache.org> wrote: > >> Thanks Zakelly and Junrui. >> >> I was actually exploring RocksDB as a state backend and I thought maybe >> Redis could offer more features as a state backend. For e.g. maybe state >> sharing between operators, geo-red of state, partitioning etc. I understand >> these are not native use cases for Flink, but maybe something that can be >> considered in future. Maybe even as an off the shelf state backend >> framework which allows embedding any other cache as a state backend. >> >> The links you shared are useful and will really help me. Really >> appreciate it. >> >> Thanks >> >> On Tuesday, 30 January, 2024 at 01:43:14 pm IST, Zakelly Lan < >> zakelly....@gmail.com> wrote: >> >> >> And I found some previous discussion, FYI: >> 1. https://issues.apache.org/jira/browse/FLINK-3035 >> 2. https://www.mail-archive.com/dev@flink.apache.org/msg10666.html >> >> Hope this helps. >> >> Best, >> Zakelly >> >> On Tue, Jan 30, 2024 at 4:08 PM Zakelly Lan <zakelly....@gmail.com> >> wrote: >> >> Hi Chirag >> >> That's an interesting idea. IIUC, storing key-values can be simply >> implemented for Redis, but supporting checkpoint and recovery is relatively >> challenging. Flink's checkpoint should be consistent among all stateful >> operators at the same time. For an *embedded* and *file-based* key value >> store like RocksDB, it is easier to implement by uploading files of >> specific time asynchronously. >> >> Moreover if you want to store your state basically in memory, then why >> not using the HashMapStateBackend. It saves the overhead of serialization >> and deserialization and may achieve better performance compared with Redis >> I guess. >> >> >> Best, >> Zakelly >> >> On Tue, Jan 30, 2024 at 2:15 PM Chirag Dewan via user < >> user@flink.apache.org> wrote: >> >> Hi, >> >> I was looking at the FLIP-254: Redis Streams Connector and I was >> wondering if Flink ever considered Redis as a state backend? And if yes, >> why was it discarded compared to RocksDB? >> >> If someone can point me towards any deep dives on why RocksDB is a better >> fit as a state backend, it would be helpful. >> >> Thanks, >> Chirag >> >>
