Undefined results when using ReadFrom with SCAN family commands #3287
Description
Bug Report
Current Behavior
We have observed that when using the ReadFrom setting with SCAN family of commands, the continuation requests (with cursor) is not guaranteed to go to the same node as the first request. Due to this behaviour, the results of SCAN family requests are undefined.
We have a master-replica Redis setup with sentinel-based discovery.
Input Code
Input Code
RedisURI redisURI = RedisURI.Builder.sentinel("sentinel-host").build();
RedisClient redisClient = RedisClient.create(redisURI);
StatefulRedisMasterReplicaConnection<byte[], byte[]> conn = MasterReplica.connect(redisClient, ByteArrayCodec.INSTANCE, redisURI);
conn.setReadFrom(ReadFrom.ANY);
byte[] key = getKey();
ScanArgs scanArgs = ScanArgs.Builder.limit(256);
ScanCursor cursor = ScanCursor.INITIAL;
MapScanCursor<byte[], byte[]> resp;
do {
resp = conn.sync().hscan(key, cursor, scanArgs);
cursor = ScanCursor.of(hscan.getCursor());
// Convert and add the results to a map
} while (!resp.isFinished());Expected behavior/code
SCAN family continuation requests need to be sticky to a Redis Node. When it is used in conjunction with ReadFrom, on every continuation request - a new node can be chosen - which seems like a bug.
Environment
- Lettuce version(s): 6.4.0.RELEASE
- Redis version: 7+
Additional Context
The same support is present for ClusterScanCursor. It maintains a nodeId and routes the next command to the same node. Any reason for not adding this in Master-replica setup?