Roshi implements a time-series event storage via a LWW-element-set CRDT with limited inline garbage collection. Roshi is a stateless, distributed layer on top of Redis and is implemented in Go. It is partition tolerant, highly available and eventually consistent.
At a high level, Roshi maintains sets of values, with each set ordered according to (external) timestamp, newest-first. Roshi provides the following API:
- Insert(key, timestamp, value)
- Delete(key, timestamp, value)
- Select(key, offset, limit) TimestampValue
Roshi stores a sharded copy of your dataset in multiple independent Redis instances, called a cluster. Roshi provides fault tolerance by duplicating clusters; multiple identical clusters, normally at least 3, form a farm. Roshi leverages CRDT semantics to ensure consistency without explicit consensus.
Roshi is basically a high-performance index for timestamped data. It's designed to sit in the critical (request) path of your application or service. The originating use case is the SoundCloud stream; see this blog post for details.
Theory and system properties
Roshi is a distributed system, for two reasons: it's made for datasets that don't fit on one machine, and it's made to be tolerant against node failure.
Next, we will explain the system design.
CRDTs (conflict-free replicated data types) are data types on which the same set of operations yields the same outcome, regardless of order of execution and duplication of operations. This allows data convergence without the need for consensus between replicas. In turn, this allows for easier implementation (no consensus protocol implementation) as well as lower latency (no wait-time for consensus).
Operations on CRDTs need to adhere to the following rules:
- Associativity (a+(b+c)=(a+b)+c), so that grouping doesn't matter.
- Commutativity (a+b=b+a), so that order of application doesn't matter.
- Idempotence (a+a=a), so that duplication doesn't matter.
Data types as well as operations have to be specifically crafted to meet these rules. CRDTs have known implementations for counters, registers, sets, graphs, and others. Roshi implements a set data type, specifically the Last Writer Wins element set (LWW-element-set).
This is an intuitive description of the LWW-element-set:
- An element is in the set, if its most-recent operation was an add.
- An element is not in the set, if its most-recent operation was a remove.
A more formal description of a LWW-element-set, as informed by Shapiro, is as follows: a set S is represented by two internal sets, the add set A and the remove set R. To add an element e to the set S, add a tuple t with the element and the current timestamp t=(e, now()) to A. To remove an element from the set S, add a tuple t with the element and the current timestamp t=(e, now()) to R. To check if an element e is in the set S, check if it is in the add set A and not in the remove set R with a higher timestamp.
Roshi implements the above definition, but extends it by applying a sort of instant garbage collection. When inserting an element E to the logical set S, check if E is already in the add set A or the remove set R. If so, check the existing timestamp. If the existing timestamp is lower than the incoming timestamp, the write succeeds: remove the existing (element, timestamp) tuple from whichever set it was found in, and add the incoming (element, timestamp) tuple to the add set A. If the existing timestamp is higher than the incoming timestamp, the write is a no-op.
Below are all possible combinations of add and remove operations. A(elements...) is the state of the add set. R(elements...) is the state of the remove set. An element is a tuple with (value, timestamp). add(element) and remove(element) are the operations.
|Original state||Operation||Resulting state|
|A(a,1) R()||add(a,0)||A(a,1) R()|
|A(a,1) R()||add(a,1)||A(a,1) R()|
|A(a,1) R()||add(a,2)||A(a,2) R()|
|A(a,1) R()||remove(a,0)||A(a,1) R()|
|A(a,1) R()||remove(a,1)||A(a,1) R()|
|A(a,1) R()||remove(a,2)||A() R(a,2)|
|A() R(a,1)||add(a,0)||A() R(a,1)|
|A() R(a,1)||add(a,1)||A() R(a,1)|
|A() R(a,1)||add(a,2)||A(a,2) R()|
|A() R(a,1)||remove(a,0)||A() R(a,1)|
|A() R(a,1)||remove(a,1)||A() R(a,1)|
|A() R(a,1)||remove(a,2)||A() R(a,2)|
For a Roshi LWW-element-set, an element will always be in either the add or the remove set exclusively, but never in both and never more than once. This means that the logical set S is the same as the add set A.
Every key in Roshi represents a set. Each set is its own LWW-element-set.
For more information on CRDTs, the following resources might be helpful:
- The chapter on CRDTs in "Distributed Systems for Fun and Profit" by Mixu
- "A comprehensive study of Convergent and Commutative Replicated Data Types" by Mark Shapiro et al. 2011
Roshi replicates data over several non-communicating clusters. A typical replication factor is 3. Roshi has two methods of replicating data: during write, and during read-repair.
A write (Insert or Delete) is sent to all clusters. The overall operation returns success the moment a user-defined number of clusters return success. Unsuccessful clusters might either have been too slow (but still accepted the write) or failed (due to a network partition or an instance crash). In case of failure, read-repair might be triggered on a later read.
A read (Select) is dependent on the read strategy employed. If the strategy queries several clusters, it might be able to spot disagreement in the returned sets. If so, the unioned set is returned to the client, and in the background, a read-repair is triggered, which lazily converges the sets across all replicas.
Package farm explains replication, read strategies, and read-repair further.
Roshi runs as a homogenous distributed system. Each Roshi instance can serve all requests (Insert, Delete, Select) for a client, and communicates with all Redis instances.
A Roshi instance is effectively stateless, but holds transient state. If a Roshi instance crashes, two types of state are lost:
- Current client connections are lost. Clients can reconnect to another Roshi instance and re-execute their operation.
- Unresolved read-repairs are lost. The read-repair might be triggered again during another read.
Since all operations are idempotent, both failure modes do not impede on convergence of the data.
Persistence is delegated to Redis. Data on a crashed-but-recovered Redis instance might be lost between the time it commited to disk, and the time it accepts connections again. The lost data gap might be repaired via read-repair.
If a Redis instance is permanently lost and has to be replaced with a fresh instance, there are two options:
- Replace it with an empty instance. Keys will be replicated to it via read-repair. As more and more keys are replicated, the read-repair load will decrease and the instance will work normally. This process might result in data loss over the lifetime of a system: if the other replicas are also lost, non-replicated keys (keys that have not been requested and thus did not trigger a read-repair) are lost.
- Replace it with a cloned replica. There will be a gap between the time of the last write respected by the replica and the first write respected by the new instance. This gap might be fixed by subsequent read-repairs.
Both processes can be expedited via a keyspace walker process. Nevertheless, these properties and procedures warrant careful consideration.
Responses to write operations
Write operations (insert or delete) return boolean to indicate whether the operation was successfully applied to the data layer, respecting the configured write quorum. Clients should interpret a write response of false to mean they should re-submit their operation. A write response of true does not imply the operation mutated the state in a way that will be visible to readers, merely that it was accepted and processed according to CRDT semantics.
As an example, all of these write operations would return true.
|Write operation||Final state||Operation description|
|Insert("foo", 3, "bar")||foo+ bar/3
|Insert("foo", 3, "bar")||foo+ bar/3
|No-op: incoming score doesn't beat existing score|
|Delete("foo", 2, "bar")||foo+ bar/3
|No-op: incoming score doesn't beat existing score|
|Delete("foo", 4, "bar")||foo+ —
|"bar" moves from add set to remove set|
|Delete("foo", 5, "bar")||foo+ —
|score of "bar" in remove set is incremented|
Roshi does not support elasticity. It is not possible to change the sharding configuration during operations. Roshi has static service discovery, configured during startup.
Roshi works with LWW-element-sets only. Clients might choose to model other data types on top of the LWW-element-sets themselves.
Correct client timestamps
Client timestamps are assumed to correctly represent the physical order of events coming into the system. Incorrect client timestamps might lead to values of a client either never appearing or always overriding other values in a set.
Assuming a replication factor of 3, and a write quorum of 2 nodes, Roshi makes the following guarantees in the presence of failures of Redis instances that represent the same data shard:
|1||No||Success dependent on read strategy||Succeed|
|2||No||Success dependent on read strategy||Fail|
Package farm explains read strategies further.
Failures of Redis instances over independent data shards don't affect instantaneous data durability. However, over time, independent Redis instance failures can lead to data loss, especially on keys which are not regularly read-repaired. In practice, a number of strategies may be used to probabilisticly mitigate this concern. For example, walking modified keys after known outages, or the whole keyspace at regular intervals, which will trigger read-repairs for inconsistent sets. However, Roshi fundamentally does not guarantee perfect data durability. Therefore, Roshi should not be used as a source of truth, but only as an intermediate store for performance critical data.
Authentication, authorization, validation
In case it's not obvious, Roshi performs no authentication, authorization, or any validation of input data. Clients must implement those things themselves.
Roshi has a layered architecture, with each layer performing a specific job with a relatively small surface area. From the bottom up...
Redis: Roshi is ultimately implemented on top of Redis instance(s), utilizing the sorted set data type. For more details on how the sorted sets are used, see package cluster, below.
Package pool performs key-based sharding over one or more Redis instances. It exposes basically a single method, taking a key and yielding a connection to the Redis instance that should hold that key. All Redis interactions go through package pool.
Package cluster implements an Insert/Select/Delete API on top of package pool. To ensure idempotency and commutativity, package cluster expects timestamps to arrive as float64s, and refuses writes with smaller timestamps than what's already been persisted. To ensure information isn't lost via deletes, package cluster maintains two physical Redis sorted sets for every logical (user) key, and manages the transition of key-timestamp-value tuples between those sets.
Package farm implements a single Insert/Select/Delete API over multiple underlying clusters. Writes (Inserts and Deletes) are sent to all clusters, and a quorum is required for success. Reads (Selects) abide one of several read strategies. Some read strategies allow for the possibility of read-repair.
roshi-walker walks the keyspace in semirandom order at a defined rate, making Select requests for each key in order to trigger read repairs.
The big picture
(Clusters need not have the same number of Redis instances.)
go build ./...
go test ./...