When a unary request takes more than predefined duration, this request
is defined as "expensive" and a warning is printed. The expensive request
duration is hard-coded to 300 ms. It can be not enough for example
for transactions with a lot of operations. The warnings just blow up
the log files and reduce throughput.
This fix allows user to configure the "expensive" request duration.
Signed-off-by: Alexey Roytman <roytman@il.ibm.com>
Before this patch, a client which cancels the context for a watch results in the
server generating a `rpctypes.ErrGRPCNoLeader` error that leads the recording of
a gRPC `Unavailable` metric in association with the client watch cancellation.
The metric looks like this:
grpc_server_handled_total{grpc_code="Unavailable",grpc_method="Watch",grpc_service="etcdserverpb.Watch",grpc_type="bidi_stream"}
So, the watch server has misidentified the error as a server error and then
propagates the mistake to metrics, leading to a false indicator that the leader
has been lost. This false signal then leads to false alerting.
The commit 9c103dd0dedfc723cd4f33b6a5e81343d8a6bae7 introduced an interceptor which wraps
watch streams requiring a leader, causing those streams to be actively canceled
when leader loss is detected.
However, the error handling code assumes all stream context cancellations are
from the interceptor. This assumption is broken when the context was canceled
because of a client stream cancelation.
The core challenge is lack of information conveyed via `context.Context` which
is shared by both the send and receive sides of the stream handling and is
subject to cancellation by all paths (including the gRPC library itself). If any
piece of the system cancels the shared context, there's no way for a context
consumer to understand who cancelled the context or why.
To solve the ambiguity of the stream interceptor code specifically, this patch
introduces a custom context struct which the interceptor uses to expose a custom
error through the context when the interceptor decides to actively cancel a
stream. Now the consuming side can more safely assume a generic context
cancellation can be propagated as a cancellation, and the server generated
leader error is preserved and propagated normally without any special inference.
When a client cancels the stream, there remains a race in the error handling
code between the send and receive goroutines whereby the underlying gRPC error
is lost in the case where the send path returns and is handled first, but this
issue can be taken separately as no matter which paths wins, we can detect a
generic cancellation.
This is a replacement of https://github.com/etcd-io/etcd/pull/11375.
Fixes#10289, #9725, #9576, #9166
