Developing Applications with Geo-Distributed CRDBs on Redis Enterprise Software (RS)
Developing geo-distributed, multi-master applications can be difficult. Application developers may have to understand a large number of race conditions between updates to various sites, network, and cluster failures that could reorder the events and change the outcome of the updates performed across geo-distributed writes.
CRDBs are geo-distributed databases that span multiple RS clusters. CRDBs depend on multi-master replication (MMR) and Conflict-free Replicated Data Types (CRDTs) to power a simple development experience for geo-distributed applications. CRDBs allow developers to use existing Redis data types and commands, but understand the developers intent and automatically handle conflicting concurrent writes to the same key across multiple geographies. For example, developers can simply use the INCR or INCRBY method in Redis in all instances of the geo-distributed application, and CRDBs handle the additive nature of INCR to reflect the correct final value. The following example displays a sequence of events over time : t1 to t9. This CRDB has two member CRDBs : member CRDB1 and member CRDB2. The local operations executing in each member CRDB is listed under the member CRDB name. The "Sync" even represent the moment where synchronization catches up to distribute all local member CRDB updates to other participating clusters and other member CRDBs.
|Time||Member CRDB1||Member CRDB2|
|t1||INCRBY key1 7|
|t2||INCRBY key1 3|
|t4||— Sync —||— Sync —|
|t6||DECRBY key1 3|
|t7||INCRBY key1 6|
|t8||— Sync —||— Sync —|
Databases provide various approaches to address some of these concerns:
- Active-Passive Geo-distributed deployments: With active-passive distributions, all writes go to an active cluster. Redis Enterprise provides a "Replica Of" capability that provides a similar approach. This can be employed when the workload is heavily balanced towards read and very few writes. However, WAN performance and availability is quite flaky and traveling large distances for writes take away from application performance and availability.
- Two-phase Commit (2PC): This approach is designed around a protocol that commits a transaction across multiple transaction managers. Two-phase commit provides a consistent transactional write across regions but fails transactions unless all participating transaction managers are "available" at the time of the transaction. The number of messages exchanged and its cross-regional availability requirement make two-phase commit unsuitable for even moderate throughputs and cross-geo writes that go over WANs.
- Sync update with Quorum-based writes: This approach synchronously coordinates a write across majority number of replicas across clusters spanning multiple regions. However, just like two-phase commit, number of messages exchanged and its cross-regional availability requirement make geo-distributed quorum writes unsuitable for moderate throughputs and cross geo writes that go over WANs.
- Last-Writer-Wins (LWW) Conflict Resolution: Some systems provide simplistic conflict resolution for all types of writes where the system clocks are used to determine the winner across conflicting writes. LWW is lightweight and can be suitable for simpler data. However, LWW can be destructive to updates that are not necessarily conflicting. For example adding a new element to a set across two geographies concurrently would result in only one of these new elements appearing in the final result with LWW.
- MVCC (multi-version concurrency control): MVCC systems maintain multiple versions of data and may expose ways for applications to resolve conflicts. Even though MVCC system can provide a flexible way to resolve conflicting writes, it comes at a cost of great complexity in the development of a solution.
Even though types and commands in CRDBs look identical to standard Redis types and commands, the underlying types in RS are enhanced to maintain more metadata to create the conflict-free data type experience. This section explains what you need to know about developing with CRDBs on Redis Enterprise Software.
CRDBs act very much like a standard Redis database except a few differences:
- CRDBs in this version support all major Redis data types. See the list of types supported in CRDBs under the Data Types section.
- As conflict handling rules differ between data types, some commands have slightly different requirements in CRDBs vs standard Redis databases. (ex: String type)
Even though the data types and methods look identical in standard Redis and CRDBs, there are specific rules that govern the handling of conflicting concurrent writes to each type.
From a developer's perspective, most supported datatypes work the same as standard Redis. However, a few methods also come with specific requirements in CRDBs.
Below is a table of the primary data types and their support levels, followed by descriptions:
|Data Type||Support Level|
|Hashes||Supported; Hash fields are treated as strings or counters|
|Bitsets||Not currently supported|
|Streams||Not currently supported|
Other Data Types
Bitmap, Bitfields, and Hyperloglog data types and operations are not currently supported in this version of CRDBs.
CRDB supports Lua scripts, but unlike standard Redis, Lua scripts always execute in effects replication mode. There is currently no way to execute them in script-replication mode.
CRDBs always operate in no eviction mode. The reasoning is that if memory is low, eviction may not help (or even worse) until garbage collection takes place.
Expiration is supported with special multi-master semantics.
If a key's expiration time is changed at the same time on different members of the CRDB, the longer extended time set via TTL on a key is preserved. As an example:
If this command was performed on key1 on cluster #1
127.0.0.1:6379> EXPIRE key1 10
And if this command was performed on key1 on cluster #2
127.0.0.1:6379> EXPIRE key1 50
The EXPIRE command setting the key to 50 would win.
And if this command was performed on key1 on cluster #3:
127.0.0.1:6379> PERSIST key1
It would win out of the three clusters hosting the CRDB as it sets the TTL on key1 to an infinite time.
The replica responsible for the "winning" expire value is also responsible to expire the key and propagate a DEL effect when this happens. A "losing" replica is from this point on not responsible for expiring the key, unless another EXPIRE command resets the TTL. Furthermore, a replica that is NOT the "owner" of the expired value:
- Silently ignores the key if a user attempts to access it in READ mode, e.g. treating it as if it was expired but not propagating a DEL.
- Expires it (sending a DEL) before making any modifications if a user attempts to access it in WRITE mode.
If a member CRDB is in an out of memory situation, that member is marked "inconsistent" by RS, the member stops responding to user traffic, and the syncer initiates full reconciliation with other peers in the CRDB.
CRDB Key Counts
Keys are counted differently for CRDBs:
- DBSIZE (in
shard-cli dbsize) reports key header instances that represent multiple potential values of a key before a replication conflict is resolved.
- expired_keys (in
bdb-cli info) can be more than the keys count in DBSIZE (in
shard-cli dbsize) because expires are not always removed when a key becomes a tombstone. A tombstone is a key that is logically deleted but still takes memory until it is collected by the garbage collector.
- The Expires average TTL (in
bdb-cli info) is computed for local expires only.
The INFO command has an additional crdt section which provides advanced troubleshooting information (applicable to support etc.):
|CRDT Context||crdt_config_version||Currently active CRDB configuration version.|
|crdt_slots||Hash slots assigned and reported by this shard.|
|crdt_replid||Unique Replica/Shard IDs.|
|crdt_clock||Clock value of local vector clock.|
|crdt_ovc||Locally observed CRDB vector clock.|
|Peers||A list of currently connected Peer Replication peers. This is similar to the slaves list reported by Redis.|
|Backlogs||A list of Peer Replication backlogs currently maintained. Typically in a full mesh topology only a single backlog is used for all peers, as the requested Ids are identical.|
|CRDT Stats||crdt_sync_full||Number of inbound full synchronization processes performed.|
|crdt_sync_partial_ok||Number of partial (backlog based) re-synchronization processes performed.|
|crdt_sync_partial-err||Number of partial re-synchronization processes failed due to exhausted backlog.|
|crdt_merge_reqs||Number of inbound merge requests processed.|
|crdt_effect_reqs||Number of inbound effect requests processed.|
|crdt_ovc_filtered_effect_reqs||Number of inbound effect requests filtered due to old vector clock.|
|crdt_gc_pending||Number of elements pending garbage collection.|
|crdt_gc_attempted||Number of attempts to garbage collect tombstones.|
|crdt_gc_collected||Number of tombstones garbaged collected successfully.|
|crdt_gc_gvc_min||The minimal globally observed vector clock, as computed locally from all received observed clocks.|
|crdt_stale_released_with_merge||Indicates last stale flag transition was a result of a complete full sync.|
|CRDT Replicas||A list of crdt_replica <uid> entries, each describes the known state of a remote instance with the following fields:|
|config_version||Last configuration version reported.|
|shards||Number of shards.|
|slots||Total number of hash slots.|
|slot_coverage||A flag indicating remote shards provide full coverage (i.e. all shards are alive).|
|max_ops_lag||Number of local operations not yet observed by the least updated remote shard|
|min_ops_lag||Number of local operations not yet observed by the most updated remote shard|