Redis LOADING error should cause different node to be queried during pipeline reads in redis-cluster #1031
Description
When processing pipeline read commands in redis-cluster if LOADING error is encountered we retry to the same node because failedCmds is not updated in pipelineReadCmds() (see https://github.com/go-redis/redis/blob/master/cluster.go#L1332). This means that if there's an extended period of time in which a node responds with LOADING error, which can happen eg after a failover recovery event in which an entire read replica needs to be replicated from the master, no pipeline reads can succeed. We unfortunately experienced this very scenario in production recently.
Is it by design that upon LOADING error is encountered during pipeline reads go-redis doesn't attempt to look for a different node to query, and if so what is/are the rationale? One factor we considered was consistency, but in redis cluster there's already no data consistency guarantee across shards anyways. Or is this an oversight that should be fixed?
As a reference for non-pipeline reads defaultProcess() already handles LOADING error properly (see https://github.com/go-redis/redis/blob/master/cluster.go#L946).