[BEAM-10259] Spanner Session leak/overload in Streaming Dataflow

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Affects Version/s: 2.18.0, 2.19.0, 2.21.0, 2.22.0
• Fix Version/s: 2.23.0
• Component/s: io-java-gcp
• Labels: dataflow gcp io
• Assignee: Niel Markwick
• Reporter: Niel Markwick
• Created: 15/Jun/20 12:13
• Updated: 28/Jun/20 08:16
• Resolved: 28/Jun/20 08:16

Description

SpannerIO.WriteToSpannerFn connects to Spanner every time @Setup is called, and closes the connection every time @Teardown is called.

This actually creates a separate Spanner connection and session pool for each WriteToSpannerFn, which generally speaking is one per thread

In single-threaded runners (eg batch dataflow on a single vCPU machine) this is not an issue, as there is normally only one WriteToSpannerFn per node/process.

In multi-threaded runners (eg streaming dataflow, or batch on multiple CPU machines), this can cause a problem with many session pools created (1 per thread) which can cause a respource leak, and is in general wasteful.

Spanner connections (and session pools) should be shared among all threads of a single process. so that the connection is only opened and closed once.

alxavier

Issue Links

• links to: GitHub Pull Request #12010