All that ApolloGateway.stop() does is clear a pollingTimer if it's currently set.

But what if pollServices is currently running (but awaiting some async operation inside updateComposition) when this is happening? Then there are two problems:

• pollServices may just end up calling pollServices again and schedule another call to itself, even though we thought we stopped it
• stop returns while stuff is still happening, which is inappropriate

I think the right fix is to have an explicit stopped flag, and also have a "pollServices is running now" promise. Something like

function pollServices() {
  // assert this.pollServicesIsRunning is null
  this.pollServicesIsRunning = actualPollServices()
    .catch(e => console.log("error in actualPollServices"));
    .then(() => {
      this.pollServicesIsRunning = null;
      if (!this.stopped) this.pollTimer = setTimeout(() => this.pollServices(), TIMEOUT);
    });
}

async function actualPollServices() {
  // ... actually poll services; no need to sleep or reschedule
}

async function stop() {
  this.stopped = true;
  if (this.pollTimer) cancelTimeout(this.pollTimer);
  if (this.pollServicesIsRunning) await this.pollServicesIsRunning;
}

With a functional ApolloGateway.stop() invoked from ApolloServer.stop(), I think we can skip the unref too.