partially full response buffers never flush
mkinsley opened this issue · 3 comments
Expect
Non-full buffers should flush in a reasonable amount of time
Observed
- Partial buffer fills on the response are not flushed to oidc-proxy client.
- From kubectl client perspective this will often show up as partial json obj... which will fail to decode.
- Pod delete will hang until something terminates the streaming connection (ie LB timeout)
- setting proxyHandler.FlushInterval = -1 resolves the issue
Symptoms
- kubectl delete pod hangs until api server connection is terminated
Fix
I set the proxy to flush immediately and the issue resolves. I'm happy to submit a PR if you would like. We could flush immediately, pick a sane value, or make it configurable.
// proxy.go: 128
// set up proxy handler using proxy
proxyHandler := httputil.NewSingleHostReverseProxy(url)
proxyHandler.Transport = p
proxyHandler.ErrorHandler = p.Error
// ----------------------------------------------------
// pod delete (anything that calls w/ watch=true ) can hang
// setting FlushInterval < 0 immediately flushes
// and solves issue
// ----------------------------------------------------
proxyHandler.FlushInterval = -1Hi @mkinsley,
It is annoying you are running into this. As you can see from the golang docs, running with a 0 value means no periodic flushing which is the current behavior.
I would like to see this changed to some sane value that can be configured. I am nervous at having a negative value as the default, perhaps 200ms would be better?
I would like more advise here on what the implications of having immediate vs periodic flushing means. In the mean time I'll put a PR together to put things in place before we agree on what the default should be.
/cc @simonswine
/assign @mkinsley