Replication and fault tolerance

Managed Service for Redis uses native Redis replication and provides high availability of cluster data using rdsync, a host status management agent.

ReplicationReplication

Managed Service for Redis clusters use asynchronous replication, i.e., the result of a write request is committed to the master host, which then forwards the data to the cluster replicas. The replication process does not affect the master availability in any way, but it can make replicas temporarily unavailable (for up to a few seconds for large databases) when loading new data into memory.

Since the replication is asynchronous, the data on replicas may be out of date: while a replica is processing updates from the master, it continues sending the existing data in response to requests, as the replica-serve-stale-data flag is set to yes.

Due to limited resources, b1, b2, and b3 class hosts are not replicated.

For more information about how replication works in Redis, read the relevant documentation.

Fault toleranceFault tolerance

To ensure fault tolerance, rdsync, a host status management agent from Yandex, was integrated into the Managed Service for Redis architecture.

Host status is stored in the distributed configuration management system. If the connection to the DCS (distributed configuration store, e.g., ZooKeeper, etcd, or Consul) is lost, the agent switches the host to protected mode and terminates client connections. When selecting a new master, if the replica host with the highest priority requires full data resync, the agent stops the replication process and selects a host with the least lag behind the master.

Thanks to the rdsync agent running in a Managed Service for Redis cluster:

• Configurations that consist of an even number of hosts (for non-sharded clusters) or one or two shards (for sharded clusters) are fault-tolerant.

• Handling client requests for the name of a host available for writes is consistent with the rdsync agent and provides up-to-date information to clients, because the statuses of all hosts are known.

• The risk of losing data is reduced if using the WAIT command with N/2 available replicas, where N is the number of cluster hosts.

Sharded clusters with the local-ssd disk type and only one host per shard are not considered fault-tolerant. You cannot create such a cluster.

Assigning a different host as a master if the primary master failsAssigning a different host as a master if the primary master fails

If the master host fails, a host with the least lag behind the master will become a new master.

You can influence master selection in a Redis cluster by configuring priorities for cluster hosts. The host with the highest priority will become a new master. If the replica host with the highest priority requires full data resync, the priority value will be ignored and a host with the least lag behind the master will become a new master.

You can set host priority:

• When creating a cluster or a host in a cluster.
• When changing the host settings.

Minimum value (lowest priority): 0. A host with such priority value can become a master only if there are no other hosts suitable for the role. Default priority value: 100. You can specify a value higher than 100.

A master host can be changed either automatically, as a result of a failure, or manually. Manual master switching is available both for sharded and unsharded clusters.

PersistencePersistence

Managed Service for Redis clusters use data persistence presets. You can disable persistence if needed to improve server throughput, as the DBMS will stop writing updates to disk.

Persistence settingsPersistence settings

By default, cluster persistence is enabled and uses the following Redis settings:

• save ""

  Regular RDB file saving is disabled. AOF mode is used instead.

• appendonly yes

  AOF (Append Only File) mode is enabled. In this mode, Redis logs every write operation without changing already written data.

• no-appendfsync-on-rewrite yes

  Since the AOF fsync policy is set to everysec, the AOF log background save process (BGSAVE) or the rewrite process (BGREWRITEAOF) performs a lot of disk I/O operations. Redis may block fsync() calls for too long in some Linux configurations.

  The setting prevents calling fsync() within the main system process when running BGSAVE or BGREWRITEAOF.

  When BGREWRITEAOF is running, fsync() is called. Redis writes the shortest sequence of commands required to restore the current dataset to memory. You can manage data size using the aof-rewrite-incremental-fsync setting.

• auto-aof-rewrite-percentage 100

  The AOF log size must exceed 100% for the AOF file to be automatically rewritten. This setting depends on the log file's auto-aof-rewrite-min-size setting.

• auto-aof-rewrite-min-size 64mb

  Auto-aof-rewrite-min-size 64mb

• aof-load-truncated yes

  Allows loading a truncated AOF file if the system crashes. The log notifies the user on loading the truncated file.

• aof-rewrite-incremental-fsync yes

  Enables AOF file synchronization every time 32 MB of data is generated.

• aof-use-rdb-preamble yes

  Enables using the RDB file as a prefix to the AOF file when rewriting or restoring it.

For more information on Redis persistence mechanisms, refer to the Redis documentation and the redis.conf config file description.

Disabling persistenceDisabling persistence

With persistence disabled, the following Redis settings take effect:

• save ""

  Regular RDB file saving is disabled.

• appendonly no

  AOF (Append Only File) mode is disabled.