go-redis could not load a slot map from a cluster whose nodes run cluster-preferred-endpoint-type unknown-endpoint. CLUSTER SLOTS then carries a null host, and the command failed with redis: nil. I sent the fix (PR #4131, merged 2026-10-26, about ten days after I opened it). The PR links no issue.
When it happens
With that setting a node sends a null host in CLUSTER SLOTS. redis.conf says the client should reuse the host it sent the request to, with the port from the reply. Null is $-1 on RESP2 and _ on RESP3. ClusterSlotsCmd reads the host with ReadString, which returns redis.Nil for both. The command failed, the rest of the reply stayed on the pooled connection, and a ClusterClient could not load its slot map. I reproduced it against a real cluster node on RESP2 and RESP3. @ndyakov checked it against a fake node, and later on a local six-node cluster with unknown-endpoint on every node, on both protocols.
Why the gap existed
The host read, before the fix:
ip, err := rd.ReadString()
if err != nil {
return err
}
Any error ended the command, Nil included. Per @ndyakov, Redis sends the null from addNodeToNodeReply in cluster.c, in the unknown-endpoint branch, with addReplyNull. The sources say nothing about why Nil was not handled here, so I stop at what the code did.
The fix
Two parts. In command.go, Nil is no longer an error for the host:
- if err != nil {
+ // A null host (cluster-preferred-endpoint-type unknown-endpoint)
+ // means "reuse the origin host"; newClusterState fills it in.
+ if err != nil && err != Nil {
In osscluster.go, newClusterState fills an empty host from the address that returned the slot map. The new function sits next to the existing port 0 substitution:
func replaceEmptyHost(nodeAddr, originHost string) string {
nodeHost, nodePort, err := net.SplitHostPort(nodeAddr)
if err != nil || nodeHost != "" {
return nodeAddr
}
return net.JoinHostPort(originHost, nodePort)
}
NodeAddress keeps the raw :port value, as it does for the other substitutions. One side effect: an empty-string host, which a node sends when it does not know its own IP, used to dial :port on the local machine and now takes the origin host, as redis-py does. An existing loopback spec changed to expect that.
The test
TestNewClusterState_NullHost loops over both encodings, _\r\n and $-1\r\n, through readReply and newClusterState. It checks the dial address, and that NodeAddress stays raw:
if got := client.NodeAddress(); got != ":6380" {
Per @ndyakov it fails on master with readReply() = redis: nil, and with the osscluster.go change reverted it fails with Addr = ":6380", so it covers both halves.
What review changed
The PR has five commits. The repo’s review bot and @ndyakov raised these points:
- The bot, P1: an existing spec still expected
:1234for an empty host. Fixed in the second commit. - The bot, P2: when the null-host entry is the origin node itself, the cached node keeps no raw
NodeAddress. I answered that this is not specific to the empty host:GetOrCreateisGetOrCreateWithNodeAddress(addr, ""), so the seed node has none. I kept it out of this PR and offered a separate one. @ndyakov agreed it was already on master. - @ndyakov, nit: say why
Nilis accepted, so the guard is not removed as a typo. That is the comment in the diff above. - The bot, P1: add a real-cluster test. I added a Ginkgo spec on the six-node cluster with
unknown-endpointon every node. It loads the state, runs SET and GET on two masters and forces a MOVED. - The bot, P1, and @ndyakov: that MOVED part passes only because the test cluster is on loopback. With
unknown-endpointthe redirect is-MOVED <slot> :<port>, and:portdials the local machine. The last commit only reworded the spec comment andIttext to say so. Using the host of the redirecting node is a follow-up.
@ndyakov’s last approval was on 2026-10-22, and they merged it on 2026-10-26.
The lesson
Nilfrom a typed read can be a valid value. Check what each non-error reply shape means before a guard returns on it.- Say why an odd guard is there, in a comment, at the guard.
- Say what a test does not cover. This one stops at loopback.
- Check the sibling command. @ndyakov found that
CLUSTER SHARDSsendsendpointas an empty bulk string, not NULL, soClusterShardsCmddoes not have this bug.
If a reader of yours returns on a null that the protocol defines as a value, reply on X.