What happened?
Currently SDKs communicating to the runner initiate a single DataStream rpc and all data corresponding to process bundle requests on the sdk receive data on this stream. This data is pushed by the runner and consumed by sdks by queueing on local queues per-bundle but these may fill up (due to size limits or perhaps improved in the future as memory limits to prevent OOMs). However blocking on a per-bundle queue means that other data on the grpc is blocked from being read. This can lead to scenarios where a single slow process bundle can fill up it's queue and then prevent other bundles from receiving data, leading to poor utilization.
A proposed fix is to allow for the use of separate "named" data streams and have the runner indicate which data stream should be used for each bundle. A bundle not consumign it's input data queue will just block the single grpc it is assigned to, and thus isolation can be achieved by assigning unrelated bundles to unique stream ids.
Note that this is more of a priority for Java SDK (and go?) because Python generally runs with an sdk per cpu and thus gets more stream isolation already.
Issue Priority
Priority: 2 (default / most bugs should be filed as P2)
Issue Components
What happened?
Currently SDKs communicating to the runner initiate a single DataStream rpc and all data corresponding to process bundle requests on the sdk receive data on this stream. This data is pushed by the runner and consumed by sdks by queueing on local queues per-bundle but these may fill up (due to size limits or perhaps improved in the future as memory limits to prevent OOMs). However blocking on a per-bundle queue means that other data on the grpc is blocked from being read. This can lead to scenarios where a single slow process bundle can fill up it's queue and then prevent other bundles from receiving data, leading to poor utilization.
A proposed fix is to allow for the use of separate "named" data streams and have the runner indicate which data stream should be used for each bundle. A bundle not consumign it's input data queue will just block the single grpc it is assigned to, and thus isolation can be achieved by assigning unrelated bundles to unique stream ids.
Note that this is more of a priority for Java SDK (and go?) because Python generally runs with an sdk per cpu and thus gets more stream isolation already.
Issue Priority
Priority: 2 (default / most bugs should be filed as P2)
Issue Components