[CASSANDRA-11839] Active streams fail with SocketTimeoutException

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Component/s: Legacy/Streaming and Messaging
• Severity: Normal
• Assignee: Paulo Motta
• Reporter: Paulo Motta

Authors

Paulo Motta

Reviewers

Yuki Morishita

• Created: 18/May/16 19:41
• Updated: 16/Apr/19 09:30
• Resolved: 31/Aug/16 21:11

Description

The original reasoning behind streaming_socket_timeout_in_ms was to kill one-sided hanging streams (CASSANDRA-3838). This was never much of a problem when the default was zero (never timeout).

On CASSANDRA-8611 we changed the default to 1 hour, but it was never enforced due to CASSANDRA-11286, which was fixed recently.

On recent releases we've been receiving reports of stream failures when streaming large files, because the sender incoming socket becomes inactive, times out after 1 hour, and the stream session fails with SocketTimeoutException (CASSANDRA-11345, CASSANDRA-11826), even though the stream session is still active. The session also fails if 2i/MV rebuild takes longer than 1 hour on the receiver (CASSANDRA-8343).

The definitive fix on trunk is to add a KeepAlive message to the stream protocol to detect broken connections and retire streaming_socket_timeout_in_ms. But we must also increase the default streaming_socket_timeout_in_ms in older versions to a more conservative value, so it is still able to detect long hanging streams.

Issue Links

• contains: CASSANDRA-8343 Secondary index creation causes moves/bootstraps to fail
• is duplicated by: CASSANDRA-11826 Repairs failing for some of bigger tables