MOD-15749 Short-form CLUSTERSET: reject when the cluster API has no usable node endpoints

[gabsow]

Problem

The short form of CLUSTERSET derives each shard's ip:port from RedisModule_GetClusterNodeInfo. On Redis builds where the module cluster topology isn't fully populated — notably Redis Enterprise without the node-port backport, where getNodeDefaultClientPort() returns 0 (RED-202230) — that API yields no usable endpoint (port 0 / empty ip), and today's code:

• doesn't validate the port/ip (only checks rc != REDISMODULE_OK), so it builds nodes pointing at port 0 — or skips all nodes and builds an empty topology;
• still return REDISMODULE_OK, so the caller (DMC) gets +OK, believes the cluster is configured, and never falls back to the long form → cross-shard commands silently break;
• RedisModule_Assert(slots != NULL) can crash the shard outright.

RediSearch hit exactly this in RE and fails loudly (returns an "invalid topology" error). LibMR currently fails silently — worse.

Fix

Mirror RediSearch's defensive handling in SetClusterDataShortForm:

1. Skip nodes with no usable endpoint (rc != OK || port <= 0 || ip[0] == '\0'); zero-init the locals so the check is reliable.
2. Replace RedisModule_Assert(slots != NULL) with a soft skip (no field crash).
3. If no valid master shards remain, MR_ClusterFree() and return an error so the caller can fall back to the long-form CLUSTERSET (which carries endpoints explicitly).

This does not make the short form work in RE (that needs the core fix, RED-202230). It makes the module fail safely there — return an error instead of silently installing a broken topology — which is a prerequisite for DMC's try-then-fallback.

Testing

cluster.c compiles and libmr.a links clean. The new paths require the RE port bug to exercise (no OSS redis returns port 0), so they're covered by code review here + the enterprise beta; the existing testShortFormClusterSetWithoutSlotRangesApi (no-API path) and the RedisTimeSeries oss-cluster flow test (success path) are unaffected.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes cluster topology installation for a critical cross-shard path; behavior is more defensive but alters success/failure when the cluster API is incomplete (e.g. enterprise without port backport).

Overview
Short-form CLUSTERSET in SetClusterDataShortForm no longer treats a partially empty Redis module cluster API as success.

Nodes are skipped when RedisModule_GetClusterNodeInfo fails or returns an empty IP or non-positive port (e.g. Redis Enterprise without the node-port backport, RED-202230). RedisModule_Assert(slots != NULL) is replaced with a warning and skip when slot ranges are missing.

If no valid master shards remain after that pass, the half-built topology is torn down via MR_ClusterFree() and the function returns REDISMODULE_ERR so callers like DMC can fall back to long-form CLUSTERSET instead of replying OK with port-0 or empty topology.

^{Reviewed by Cursor Bugbot for commit 594375d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[gabsow]

Superseded by #103, which combines this with the sibling short-form CLUSTERSET fail-safe into a single PR (one review + one pin bump). Closing in favour of #103.