Open chewbranca opened 8 months ago
good idea, but let's protect ourselves from client input. rejecting an x-couch-request-id that's above a certain length, say.
+1 to check the length and maybe also ensure it has only alphanumeric characters, no escapes, slashes, semicolons, etc.
agreed on those extra checks, nick.
and a 400 bad request if those checks aren't met, rather than silently ignoring the header or stripping it of invalid chars, etc.
Summary
As mentioned in https://github.com/apache/couchdb/pull/4990, when using a load balancer or proxy in front of CouchDB, the proxy will not receive the chosen
nonce
value from CouchDB until the response headers are sent, which, in the event of long running_find
queries that return no data, can be a considerable amount of time longer than typical proxy timeouts.The result is that the Mango reports are logged with the
nonce
value but the proxy timed out the request before getting any response headers so it was never able to retrieve thenonce
for logging and connecting the proxy logs with the mango report. This is exacerbated bychttpd:maybe_log
typically being set to false when using a proxy alongside a CouchDB cluster.I was going to suggest modifying
chttpd:maybe_log
to be configurable based on error types, but @rnewson had a much simpler suggestion of allowing the proxy to provide thenonce
, thereby establishing the connection from the get go.I suggest we stick with the existing naming and allow for a request header named
X-Couch-Request-ID
, following https://github.com/apache/couchdb/blob/main/src/chttpd/src/chttpd.erl#L1368.Desired Behaviour
A frontend proxy load balancer to a CouchDB cluster may supply an
X-Couch-Request-ID
to utilize as thenonce
value instead of randomly generating one here: https://github.com/apache/couchdb/blob/main/src/chttpd/src/chttpd.erl#L297-L317Possible Solution
Additional context
https://github.com/apache/couchdb/pull/4990