Implement per-user rate limits #251
Conversation
This is the simplest version for now, just based on flag-configured rate limits per distributor process and no knowledge about total (cross-distributor rates) yet. Fixes #128
|
The |
| @@ -320,6 +338,19 @@ func (d *Distributor) Push(ctx context.Context, req *remote.WriteRequest) (*cort | |||
| return &cortex.WriteResponse{}, nil | |||
| } | |||
|
|
|||
| func (d *Distributor) getOrCreateIngestLimiter(userID string) *rate.Limiter { | |||
| d.ingestLimitersMtx.Lock() | |||
There was a problem hiding this comment.
Note I'm intentionally just going for a normal mutex here because I figured an RWLock didn't really matter in this path (access once per HTTP request).
There was a problem hiding this comment.
I'd have probably gone with a RWMutex, I expecting each ingester to be able to handle ~1million samples/s, or 10k http requests a second. Right now they're only doing ~100 req/s, so its not a big deal. We can change later.
|
I see https://godoc.org/golang.org/x/time/rate post-dates our own token bucket implementation in Weave Net that we did in weaveworks/weave#1124. |
|
@rade Ah, didn't know about Weave's implemenation. Anyways, this one here seems to be the current Go way to do it (plus the Weave one doesn't have a non-blocking method yet). Could make sense to switch the other one too. |
Indeed. That's why I mentioned it :) |
|
You've pre-guessed my only comment, so LGTM! |
This is the simplest version for now, just based on flag-configured
rate limits per distributor process and no knowledge about total
(cross-distributor rates) yet.
Fixes #128