etcdserver: requested lease not found

[sh1ng]

Bug reporting

We use an etcd cluster of 3 members with version
quay.io/coreos/etcd:v3.2.0
running in k8s cluster. Not a heavy load, about 10k writes a day.

lease, err := client.Lease.Grant(myContext, int64(24 * time.Hour))
if err != nil {
	log.Error("Unable to get a lease from etcd", "error", err)
	return err
}
_, err = client.KV.Put(myContext, key, string(msg), clientv3.WithLease(lease.ID))
if err != nil {
	log.Error("Unable to put a record into etcd", "error", err) // <- it's logged here 
	return err
}

How that even possible? We see the error a few times a day.
What is the best way to deal with it?

[gyuho]

Can you provide reproducible steps (locally)?

[sh1ng]

Not really, it's just a simple service that accepts records by grpc and store them in etcd for 24 hrs.

If you can give an idiomatic example for the scenario it would be nice.

[sh1ng]

And the main question how that could be possible that successfully created lease can't be found.

[gyuho]

@sh1ng Do you have etcd server logs when this happened?

[heyitsanthony]

@sh1ng etcd:v3.2.0? there was a restore bug fixed by 4526284; try upgrading to latest 3.2

[sh1ng]

Even after upgrade on quay.io/coreos/etcd:v3.2.16 we still see the same error.

Could it be possible that when parent context has been canceled etcd(or etcd client) returns revoked lease? We use streaming api to our service and a client might cancel it from time to time.

[gyuho]

@sh1ng Do you server logs when this happened?

[yudai]

int64(24 * time.Hour) is 86400000000000 and I think what you want here is 60 * 60 * 24 = 86400, because Grant() expects a TTL in seconds.

When a lease is created with 86400000000000, it seems etcd loses the lease immediately. Probably it's just by an overflow: https://github.com/coreos/etcd/blob/master/lease/lessor.go#L599

We may want maxLeaseTTL to avoid unexpected behavior by overflows.

[gyuho]

@yudai Got it right.

It gets overflow, and sets negative expiry time value, thus lease expires before that put request.

@yudai Do you want to send a fix?

[yudai]

@gyuho What value would you suggest for the max TTL? 10 years?
or, something like (Math.MaxInt64 / time.Second) - someBuffer ?

[gyuho]

We usually lease promote with election timeout (which is 1 second by default), so (Math.MaxInt64 / time.Second) - time.Minute to be safe?

[yudai]

@gyuho thanks for the suggestion.
I chose 9,000,000,000 to make it very safe (223,372,036 seconds buffer) and easier to document/remember.

[gyuho]

The fix will be released in 3.2 and 3.3.

[sh1ng]

Thanks guys!
Now it works like a charm.