Middleware to reject requests if we have too many requests in-flight already.
The client must opt-in to this behaviour by sending us a If-Not-Overloaded: 1 header.
We expect clients to eventually give up and just retry without the header, skipping the limiting. This implies your application should have other concurrency limiting or queuing in place, if appropriate. This might just be the node event loop.
There's a wrapper for request-promise-native available in requestWorkerRuns. This
will attempt to locate a worker which isn't overloaded, and send the request to it.
If the worker doesn't support this protocol, or a free worker cannot be found,
then the request will be run anyway; the same as if request was used directly.
import { makeOverloadLimiter } from 'unless-overloaded';
const unlessOverloaded = makeOverloadLimiter(config.maxConcurrentRequests);
router.get('/unprotected', handleUnprotected);
router.get('/expensive', unlessOverloaded, handleProtected);This allows a client which retries to pick an idle backend behind a dumb,
uncooperative load-balancer, such as a Kubernetes Service.
The Service always operates in round-robin mode. If the pod that it "picks"
is overloaded, we can inform cooperative clients that they might want to try
again, and hopefully end up being served by a different Pod, which isn't
overloaded. If they retry for long enough, they may even hit a newly scheduled
Pod in the HPA.
This is an interim step between what we have now (forkbombs) and actual decoupling.
We're undecided if this is:
- a service mesh-style "fail closed" circuit-breaker (??), c.f. https://istio.io/docs/concepts/traffic-management/#circuit-breakers
- what readiness in kubernetes is supposed to be, without the alerting on failure (again, only because we're expecting the client to fail-closed)
- a work-stealing queue, but entirely upside down? A work-sliding queue. The insertion work is done by the client's probing loop; it's trying to probe for a free queue slot in various workers' queues. If it fails, it can just push it on to a random worker's queue. It's not quite work stealing because once it's actually inserted, it's stuck forever. Some work-stealing impls do this anyway.