Details
-
Improvement
-
Status: Resolved
-
Normal
-
Resolution: Fixed
-
None
-
None
Description
A serious problem is that there is no co-ordination whatsoever between gossip and the consumers of gossip. In particular, on a large cluster with hundreds of nodes, it takes several seconds for gossip to settle because the gossip stage is CPU bound. This leads to a node starting up and accessing thrift traffic long before it has any clue of what up and down. This leads to client-visible timeouts (for nodes that are down but not identified as such) and UnavailableException (for nodes that are up but not yet identified as such). This is really bad in general, but in particular for clients doing non-idempotent writes (counter increments).
I was going to fix this as part of more significant re-writing in other tickets having to do with gossip/topology/etc, but that's not going to happen. So, the attached patch is roughly what we're running with in production now to make restarts bearable. The minimum wait time is both for ensuring that gossip has time to start becoming CPU bound if it will be, and the reason it's large is to allow for down nodes to be identified as such in most typical cases with a default phi conviction threshold (untested, we actually ran with a smaller number of 5 seconds minimum, but from past experience I believe 15 seconds is enough).
The patch is tested on our 1.1 branch. It applies on trunk, and the diff is against trunk, but I have not tested it against trunk.
Attachments
Attachments
Issue Links
- is duplicated by
-
-
- Resolved
-
- is related to
-
CASSANDRA-6127 vnodes don't scale to hundreds of nodes
-
- Resolved
-
-
-
- Resolved
-