Envoy can act as a Redis proxy, partitioning commands among instances in a cluster. In this mode, the goals of Envoy are to maintain availability and partition tolerance over consistency. This is the key point when comparing Envoy to Redis Cluster. Envoy is designed as a best-effort cache, meaning that it will not try to reconcile inconsistent data or keep a globally consistent view of cluster membership.
The Redis project offers a thorough reference on partitioning as it relates to Redis. See “Partitioning: how to split data among multiple Redis instances”.
Features of Envoy Redis:
- Redis protocol codec.
- Hash-based partitioning.
- Ketama distribution.
- Detailed command statistics.
- Active and passive healthchecking.
Planned future enhancements:
- Additional timing stats.
- Circuit breaking.
- Request collapsing for fragmented commands.
- Built-in retry.
- Hash tagging.
For filter configuration details, see the Redis proxy filter configuration reference.
The corresponding cluster definition should be configured with ring hash load balancing.
If active health checking is desired, the cluster should be configured with a custom health check which configured as a Redis health checker.
If passive healthchecking is desired, also configure outlier detection.
For the purposes of passive healthchecking, connect timeouts, command timeouts, and connection close map to 5xx. All other responses from Redis are counted as a success.
At the protocol level, pipelines are supported. MULTI (transaction block) is not. Use pipelining wherever possible for the best performance.
At the command level, Envoy only supports commands that can be reliably hashed to a server. PING is the only exception, which Envoy responds to immediately with PONG. Arguments to PING are not allowed. All other supported commands must contain a key. Supported commands are functionally identical to the original Redis command except possibly in failure scenarios.
For details on each command’s usage see the official Redis command reference.
If Redis throws an error, we pass that error along as the response to the command. Envoy treats a response from Redis with the error datatype as a normal response and passes it through to the caller.
Envoy can also generate its own errors in response to the client.
|no upstream host||The ring hash load balancer did not have a healthy host available at the ring position chosen for the key.|
|upstream failure||The backend did not respond within the timeout period or closed the connection.|
|invalid request||Command was rejected by the first stage of the command splitter due to datatype or length.|
|unsupported command||The command was not recognized by Envoy and therefore cannot be serviced because it cannot be hashed to a backend server.|
|finished with n errors||Fragmented commands which sum the response (e.g. DEL) will return the total number of errors received if any were received.|
|upstream protocol error||A fragmented command received an unexpected datatype or a backend responded with a response that not conform to the Redis protocol.|
|wrong number of arguments for command||Certain commands check in Envoy that the number of arguments is correct.|
In the case of MGET, each individual key that cannot be fetched will generate an error response. For example, if we fetch five keys and two of the keys’ backends time out, we would get an error response for each in place of the value.
$ redis-cli MGET a b c d e 1) "alpha" 2) "bravo" 3) (error) upstream failure 4) (error) upstream failure 5) "echo"