This patch did give some better results, but break watches on
the services of a node.
It is possible to apply the same optimization for nodes than
to services (one index per instance), but it would complicate
further the patch.
Let's do it in another PR.
The root cause is actually that the agent's streaming HTTP API didn't flush until the first log line was found which commonly was pretty soon since the default level is INFO. In cases where there were no logs immediately due to level for instance, the client gets stuck in the HTTP code waiting on a response packet from the server before we enter the loop that checks the shutdown channel from the signal handler.
This fix flushes the initial status immediately on the streaming endpoint which lets the client code get into it's expected state where it's listening for shutdown or log lines.
This patch improves the watches for services on large cluster:
each service has now its own index, such watches on a specific service
are not modified by changes in the global catalog.
It should improve a lot the performance of tools such as consul-template
or libraries performing watches on very large clusters with many
services/watches.
- register endpoints with supported methods
- support OPTIONS requests, indicating supported methods
- extract method validation (error 405) from individual endpoints
- on 405 where multiple methods are allowed, create a single Allow
header with comma-separated values, not multiple Allow headers.
Because this code was doing pointer equality checks, it would work for
the case of a failed attempted RPC because the objects are from the
manager itself:
https://github.com/hashicorp/consul/blob/v1.0.3/agent/consul/rpc.go#L283-L302
But the pointer check would always fail for events coming in from the
Serf path because the server object is newly-created:
https://github.com/hashicorp/consul/blob/v1.0.3/agent/router/serf_adapter.go#L14-L40
This means that we didn't proactively shift RPC traffic away from a
failed server, we'd have to wait for an RPC to fail, which exposes
the error to the calling client.
By switching over to a name check vs. a pointer check we get the correct
behavior. We added a DEBUG log as well to help observe this behavior during
integrated testing.
Related to #3863 since the fix here needed the same logic duplicated, owing
to the complicated atomic stuff.
/cc @dadgar for a heads up in case this also affects Nomad.
Preetha Appan
Pierre Souchay
Mitchell Hashimoto
Edd Steel
Paul Banks
Alex Dadgar
Kyle Havlovitz
Andrei Burd
James Phillips
Wim
Veselkov Konstantin