Global memory statistician for small memory fragments

[wcy-fdu]

RisingWave will occupy some small memory during its working process, for example:

• actor channel will buffer 2048 rows(the upper limit may be size based 1MB late)
• stateful executors' mem tables will occupy memory until it's size exceed 4MB

Although these memory footprints are small, they vary dynamically with workload. If there are many materialize views/Actors in the current system, the total amount of these small memories will also increase accordingly, thereby increasing the risk of OOM(We have witnessed such issues during longevity test).

The current strategy involves global memory management, where we use jemalloc to monitor the memory usage of Compute Node. Once a certain threshold is reached, we start evicting the LRU cache. However, we currently do not track these small memories because we consider them to be small and quickly released. Since customers create a lot of MVs in the cluster, we should count these small memories and make some mitigation strategies after they add up.

possible methods:

• introduce some kind of memory tracking to monitor these fragmented memories.
• more reasonable memory allocation according to the memory usage(eg. more reserved memory or allocate a fixed size to it in streaming memory)
• introduce some kind of control strategy when these memories are too large.

[hzxa21]

RisingWave will occupy some small memory during its working process, for example:

• actor channel will buffer 2048 rows(the upper limit may be size based 1MB late)
• stateful executors' mem tables will occupy memory until it's size exceed 4MB

Two more examples worth mentioning:

• source reader may buffer source messages. For example, by per librdkafka consumer instance can buffer 64MB message by default
• sink may buffer stream chunks (without spill) per epoch in order to do compaction if sink has different PK than upstream

[fuyufjh]

Any specific tasks?

[github-actions]

This issue has been open for 60 days with no activity. Could you please update the status? Feel free to continue discussion or close as not planned.