Dynamo

Apache Cassandra relies on a number of techniques from Amazon’s Dynamo distributed storage key-value system. Each node in the Dynamo system has three main components:

• Request coordination over a partitioned dataset

• Ring membership and failure detection

• A local persistence (storage) engine

Cassandra primarily draws from the first two clustering components, while using a storage engine based on a Log Structured Merge Tree (LSM). In particular, Cassandra relies on Dynamo style:

• Dataset partitioning using consistent hashing

• Multi-master replication using versioned data and tunable consistency

• Distributed cluster membership and failure detection via a gossip protocol

• Incremental scale-out on commodity hardware

Cassandra was designed this way to meet large-scale (PiB+) business-critical storage requirements. In particular, as applications demanded full global replication of petabyte-scale datasets along with always available low-latency reads and writes, it became imperative to design a new kind of database model as the relational database systems of the time struggled to meet the new requirements of global-scale applications.

Dataset Partitioning: Consistent Hashing

Cassandra achieves horizontal scalability by partitioning) all data stored in the system using a hash function. Each partition is replicated to multiple physical nodes, often across failure domains such as racks and even datacenters. As every replica can independently accept mutations to every key that it owns, every key must be versioned. Unlike in the original Dynamo paper where deterministic versions and vector clocks were used to reconcile concurrent updates to a key, Cassandra uses a simpler last-write-wins model where every mutation is timestamped (including deletes) and then the latest version of data is the “winning” value. Formally speaking, Cassandra uses a Last-Write-Wins Element-Set conflict-free replicated data type for each CQL row, or LWW-Element-Set CRDT), to resolve conflicting mutations on replica sets.

Consistent Hashing using a Token Ring

Cassandra partitions data over storage nodes using a special form of hashing called consistent hashing. In naive data hashing, you typically allocate keys to buckets by taking a hash of the key modulo the number of buckets. For example, if you want to distribute data to 100 nodes using naive hashing you might assign every node to a bucket between 0 and 100, hash the input key modulo 100, and store the data on the associated bucket. In this naive scheme, however, adding a single node might invalidate almost all of the mappings.

Cassandra instead maps every node to one or more tokens on a continuous hash ring, and defines ownership by hashing a key onto the ring and then “walking” the ring in one direction, similar to the Chord algorithm. The main difference of consistent hashing to naive data hashing is that when the number of nodes (buckets) to hash into changes, consistent hashing only has to move a small fraction of the keys.

For example, if we have an eight node cluster with evenly spaced tokens, and a replication factor (RF) of 3, then to find the owning nodes for a key we first hash that key to generate a token (which is just the hash of the key), and then we “walk” the ring in a clockwise fashion until we encounter three distinct nodes, at which point we have found all the replicas of that key. This example of an eight node cluster with gRF=3 can be visualized as follows:

You can see that in a Dynamo-like system, ranges of keys, also known as token ranges, map to the same physical set of nodes. In this example, all keys that fall in the token range excluding token 1 and including token 2 (grange(t1, t2]) are stored on nodes 2, 3 and 4.

Multiple Tokens per Physical Node (vnodes)

Simple single-token consistent hashing works well if you have many physical nodes to spread data over, but with evenly spaced tokens and a small number of physical nodes, incremental scaling (adding just a few nodes of capacity) is difficult because there are no token selections for new nodes that can leave the ring balanced. Cassandra seeks to avoid token imbalance because uneven token ranges lead to uneven request load. For example, in the previous example there is no way to add a ninth token without causing imbalance; instead we would have to insert 8 tokens in the midpoints of the existing ranges.

The Dynamo paper advocates for the use of “virtual nodes” to solve this imbalance problem. Virtual nodes solve the problem by assigning multiple tokens in the token ring to each physical node. By allowing a single physical node to take multiple positions in the ring, we can make small clusters look larger and therefore even with a single physical node addition we can make it look like we added many more nodes, effectively taking many smaller pieces of data from more ring neighbors when we add even a single node.

Cassandra introduces some nomenclature to handle these concepts:

• Token: A single position on the Dynamo-style hash ring.

• Endpoint: A single physical IP and port on the network.

• Host ID: A unique identifier for a single “physical” node, usually present at one gEndpoint and containing one or more gTokens.

• Virtual Node (or vnode): A gToken on the hash ring owned by the same physical node, one with the same gHost ID.

The mapping of Tokens to Endpoints gives rise to the Token Map where Cassandra keeps track of what ring positions map to which physical endpoints. For example, in the following figure we can represent an eight node cluster using only four physical nodes by assigning two tokens to every node:

Multiple tokens per physical node provide the following benefits:

1. When a new node is added it accepts approximately equal amounts of data from other nodes in the ring, resulting in equal distribution of data across the cluster.

2. When a node is decommissioned, it loses data roughly equally to other members of the ring, again keeping equal distribution of data across the cluster.

3. If a node becomes unavailable, query load (especially token-aware query load), is evenly distributed across many other nodes.

Multiple tokens, however, can also have disadvantages:

1. Every token introduces up to 2 * (RF - 1) additional neighbors on the token ring, which means that there are more combinations of node failures where we lose availability for a portion of the token ring. The more tokens you have, the higher the probability of an outage.

2. Cluster-wide maintenance operations are often slowed. For example, as the number of tokens per node is increased, the number of discrete repair operations the cluster must do also increases.

3. Performance of operations that span token ranges could be affected.

Note that in Cassandra 2.x, the only token allocation algorithm available was picking random tokens, which meant that to keep balance the default number of tokens per node had to be quite high, at 256. This had the effect of coupling many physical endpoints together, increasing the risk of unavailability. That is why in 3.x + a new deterministic token allocator was added which intelligently picks tokens such that the ring is optimally balanced while requiring a much lower number of tokens per physical node.

Multi-master Replication: Versioned Data and Tunable Consistency

Cassandra replicates every partition of data to many nodes across the cluster to maintain high availability and durability. When a mutation occurs, the coordinator hashes the partition key to determine the token range the data belongs to and then replicates the mutation to the replicas of that data according to the Replication Strategy.

All replication strategies have the notion of a replication factor (RF), which indicates to Cassandra how many copies of the partition should exist. For example with a RF=3 keyspace, the data will be written to three distinct replicas. Replicas are always chosen such that they are distinct physical nodes which is achieved by skipping virtual nodes if needed. Replication strategies may also choose to skip nodes present in the same failure domain such as racks or datacenters so that Cassandra clusters can tolerate failures of whole racks and even datacenters of nodes.

Replication Strategy

Cassandra supports pluggable replication strategies, which determine which physical nodes act as replicas for a given token range. Every keyspace of data has its own replication strategy. All production deployments should use the NetworkTopologyStrategy while the SimpleStrategy replication strategy is useful only for testing clusters where you do not yet know the datacenter layout of the cluster.

NetworkTopologyStrategy

NetworkTopologyStrategy requires a specified replication factor for each datacenter in the cluster. Even if your cluster only uses a single datacenter, NetworkTopologyStrategy is recommended over SimpleStrategy to make it easier to add new physical or virtual datacenters to the cluster later, if required.

In addition to allowing the replication factor to be specified individually by datacenter, NetworkTopologyStrategy also attempts to choose replicas within a datacenter from different racks as specified by the Snitch. If the number of racks is greater than or equal to the replication factor for the datacenter, each replica is guaranteed to be chosen from a different rack. Otherwise, each rack will hold at least one replica, but some racks may hold more than one. Note that this rack-aware behavior has some potentially surprising implications. For example, if there are not an even number of nodes in each rack, the data load on the smallest rack may be much higher. Similarly, if a single node is bootstrapped into a brand new rack, it will be considered a replica for the entire ring. For this reason, many operators choose to configure all nodes in a single availability zone or similar failure domain as a single “rack”.

SimpleStrategy

SimpleStrategy allows a single integer replication_factor to be defined. This determines the number of nodes that should contain a copy of each row. For example, if replication_factor is 3, then three different nodes should store a copy of each row.

SimpleStrategy treats all nodes identically, ignoring any configured datacenters or racks. To determine the replicas for a token range, Cassandra iterates through the tokens in the ring, starting with the token range of interest. For each token, it checks whether the owning node has been added to the set of replicas, and if it has not, it is added to the set. This process continues until replication_factor distinct nodes have been added to the set of replicas.

Transient Replication

Transient replication is an experimental feature in Cassandra 4.0 not present in the original Dynamo paper. This feature allows configuration of a subset of replicas to replicate only data that hasn’t been incrementally repaired. This configuration decouples data redundancy from availability. For instance, if you have a keyspace replicated at RF=3, and alter it to RF=5 with two transient replicas, you go from tolerating one failed replica to tolerating two, without corresponding increase in storage usage. Now, three nodes will replicate all the data for a given token range, and the other two will only replicate data that hasn’t been incrementally repaired.

To use transient replication, first enable the option in cassandra.yaml. Once enabled, both SimpleStrategy and NetworkTopologyStrategy can be configured to transiently replicate data. Configure it by specifying replication factor as <total_replicas>/<transient_replicas Both SimpleStrategy and NetworkTopologyStrategy support configuring transient replication.

Transiently replicated keyspaces only support tables created with read_repair set to NONE; monotonic reads are not currently supported. You also can’t use LWT, logged batches, or counters in 4.0. You will possibly never be able to use materialized views with transiently replicated keyspaces and probably never be able to use secondary indices with them.

Transient replication is an experimental feature that is not ready for production use. The expected audience is experienced users of Cassandra capable of fully validating a deployment of their particular application. That means you have the experience to check that operations like reads, writes, decommission, remove, rebuild, repair, and replace all work with your queries, data, configuration, operational practices, and availability requirements.

Anticipated additional features in 4.next are support for monotonic reads with transient replication, as well as LWT, logged batches, and counters.

Data Versioning

Cassandra uses mutation timestamp versioning to guarantee eventual consistency of data. Specifically all mutations that enter the system do so with a timestamp provided either from a client clock or, absent a client-provided timestamp, from the coordinator node’s clock. Updates resolve according to the conflict resolution rule of last write wins. Cassandra’s correctness does depend on these clocks, so make sure a proper time synchronization process is running such as NTP.

Cassandra applies separate mutation timestamps to every column of every row within a CQL partition. Rows are guaranteed to be unique by primary key, and each column in a row resolves concurrent mutations according to last-write-wins conflict resolution. This means that updates to different primary keys within a partition can actually resolve without conflict! Furthermore the CQL collection types such as maps and sets use this same conflict-free mechanism, meaning that concurrent updates to maps and sets are guaranteed to resolve as well.

Replica Synchronization

As replicas in Cassandra can accept mutations independently, it is possible for some replicas to have newer data than others. Cassandra has many best-effort techniques to drive convergence of replicas including Replica read repair <read-repair> in the read path and Hinted handoff <hints> in the write path.

These techniques are only best-effort, however, and to guarantee eventual consistency Cassandra implements anti-entropy repair <repair> where replicas calculate hierarchical hash trees over their datasets called Merkle trees that can then be compared across replicas to identify mismatched data. Like the original Dynamo paper Cassandra supports full repairs where replicas hash their entire dataset, create Merkle trees, send them to each other and sync any ranges that don’t match.

Unlike the original Dynamo paper, Cassandra also implements sub-range repair and incremental repair. Sub-range repair allows Cassandra to increase the resolution of the hash trees (potentially down to the single partition level) by creating a larger number of trees that span only a portion of the data range. Incremental repair allows Cassandra to only repair the partitions that have changed since the last repair.

Tunable Consistency

Cassandra supports a per-operation tradeoff between consistency and availability through Consistency Levels. Cassandra’s consistency levels are a version of Dynamo’s R + W > N consistency mechanism where operators could configure the number of nodes that must participate in reads (R) and writes (W) to be larger than the replication factor (N). In Cassandra, you instead choose from a menu of common consistency levels which allow the operator to pick R and W behavior without knowing the replication factor. Generally writes will be visible to subsequent reads when the read consistency level contains enough nodes to guarantee a quorum intersection with the write consistency level.

The following consistency levels are available:

ONE

Only a single replica must respond.

TWO

Two replicas must respond.

THREE

Three replicas must respond.

QUORUM

A majority (n/2 + 1) of the replicas must respond.

ALL

All of the replicas must respond.

LOCAL_QUORUM

A majority of the replicas in the local datacenter (whichever datacenter the coordinator is in) must respond.

EACH_QUORUM

A majority of the replicas in each datacenter must respond.

LOCAL_ONE

Only a single replica must respond. In a multi-datacenter cluster, this also guarantees that read requests are not sent to replicas in a remote datacenter.

ANY

A single replica may respond, or the coordinator may store a hint. If a hint is stored, the coordinator will later attempt to replay the hint and deliver the mutation to the replicas. This consistency level is only accepted for write operations.

Write operations are always sent to all replicas, regardless of consistency level. The consistency level simply controls how many responses the coordinator waits for before responding to the client.

For read operations, the coordinator generally only issues read commands to enough replicas to satisfy the consistency level. The one exception to this is when speculative retry may issue a redundant read request to an extra replica if the original replicas have not responded within a specified time window.

Picking Consistency Levels

It is common to pick read and write consistency levels such that the replica sets overlap, resulting in all acknowledged writes being visible to subsequent reads. This is typically expressed in the same terms Dynamo does, in that W + R > RF, where W is the write consistency level, R is the read consistency level, and RF is the replication factor. For example, if RF = 3, a QUORUM request will require responses from at least 2/3 replicas. If QUORUM is used for both writes and reads, at least one of the replicas is guaranteed to participate in both the write and the read request, which in turn guarantees that the quorums will overlap and the write will be visible to the read.

In a multi-datacenter environment, LOCAL_QUORUM can be used to provide a weaker but still useful guarantee: reads are guaranteed to see the latest write from within the same datacenter. This is often sufficient as clients homed to a single datacenter will read their own writes.

If this type of strong consistency isn’t required, lower consistency levels like LOCAL_ONE or ONE may be used to improve throughput, latency, and availability. With replication spanning multiple datacenters, LOCAL_ONE is typically less available than ONE but is faster as a rule. Indeed ONE will succeed if a single replica is available in any datacenter.

Distributed Cluster Membership and Failure Detection

The replication protocols and dataset partitioning rely on knowing which nodes are alive and dead in the cluster so that write and read operations can be optimally routed. In Cassandra liveness information is shared in a distributed fashion through a failure detection mechanism based on a gossip protocol.

Gossip

Gossip is how Cassandra propagates basic cluster bootstrapping information such as endpoint membership and internode network protocol versions. In Cassandra’s gossip system, nodes exchange state information not only about themselves but also about other nodes they know about. This information is versioned with a vector clock of (generation, version) tuples, where the generation is a monotonic timestamp and version is a logical clock that increments roughly every second. These logical clocks allow Cassandra gossip to ignore old versions of cluster state just by inspecting the logical clocks presented with gossip messages.

Every node in the Cassandra cluster runs the gossip task independently and periodically. Every second, every node in the cluster:

1. Updates the local node’s heartbeat state (the version) and constructs the node’s local view of the cluster gossip endpoint state.

2. Picks a random other node in the cluster to exchange gossip endpoint state with.

3. Probabilistically attempts to gossip with any unreachable nodes (if one exists)

4. Gossips with a seed node if that didn’t happen in step 2.

When an operator first bootstraps a Cassandra cluster, they designate certain nodes as seed nodes. Any node can be a seed node, and the only difference between seed and non-seed nodes is that seed nodes are allowed to bootstrap into the ring without seeing any other seed nodes. Furthermore, once a cluster is bootstrapped, seed nodes become hotspots for gossip due to step 4 above.

As non-seed nodes must be able to contact at least one seed node in order to bootstrap into the cluster, it is common to include multiple seed nodes, often one for each rack or datacenter. Seed nodes are often chosen using existing off-the-shelf service discovery mechanisms.

Currently, gossip also propagates token metadata and schema version information. This information forms the control plane for scheduling data movements and schema pulls. For example, if a node sees a mismatch in schema version in gossip state, it will schedule a schema sync task with the other nodes. As token information propagates via gossip it is also the control plane for teaching nodes which endpoints own what data.

Ring Membership and Failure Detection

Gossip forms the basis of ring membership, but the failure detector ultimately makes decisions about if nodes are UP or DOWN. Every node in Cassandra runs a variant of the Phi Accrual Failure Detector, in which every node is constantly making an independent decision of if their peer nodes are available or not. This decision is primarily based on received heartbeat state. For example, if a node does not see an increasing heartbeat from a node for a certain amount of time, the failure detector “convicts” that node, at which point Cassandra will stop routing reads to it (writes will typically be written to hints). If/when the node starts heartbeating again, Cassandra will try to reach out and connect, and if it can open communication channels it will mark that node as available.

Cassandra will never remove a node from gossip state without explicit instruction from an operator via a decommission operation or a new node bootstrapping with a replace_address_first_boot option. This choice is intentional to allow Cassandra nodes to temporarily fail without causing data to needlessly re-balance. This also helps to prevent simultaneous range movements, where multiple replicas of a token range are moving at the same time, which can violate monotonic consistency and can even cause data loss.

Incremental Scale-out on Commodity Hardware

Cassandra scales-out to meet the requirements of growth in data size and request rates. Scaling-out means adding additional nodes to the ring, and every additional node brings linear improvements in compute and storage. In contrast, scaling-up implies adding more capacity to the existing database nodes. Cassandra is also capable of scale-up, and in certain environments it may be preferable depending on the deployment. Cassandra gives operators the flexibility to choose either scale-out or scale-up.

One key aspect of Dynamo that Cassandra follows is to attempt to run on commodity hardware, and many engineering choices are made under this assumption. For example, Cassandra assumes nodes can fail at any time, auto-tunes to make the best use of CPU and memory resources available and makes heavy use of advanced compression and caching techniques to get the most storage out of limited memory and storage capabilities.

Simple Query Model

Cassandra, like Dynamo, chooses not to provide cross-partition transactions that are common in SQL Relational Database Management Systems (RDBMS). This both gives the programmer a simpler read and write API, and allows Cassandra to more easily scale horizontally since multi-partition transactions spanning multiple nodes are notoriously difficult to implement and typically very latent.

Instead, Cassandra chooses to offer fast, consistent, latency at any scale for single partition operations, allowing retrieval of entire partitions or only subsets of partitions based on primary key filters. Furthermore, Cassandra does support single partition compare and swap functionality via the lightweight transaction CQL API.

Simple Interface for Storing Records

Cassandra, in a slight departure from Dynamo, chooses a storage interface that is more sophisticated than “simple key-value” stores but significantly less complex than SQL relational data models. Cassandra presents a wide-column store interface, where partitions of data contain multiple rows, each of which contains a flexible set of individually typed columns. Every row is uniquely identified by the partition key and one or more clustering keys, and every row can have as many columns as needed.

This allows users to flexibly add new columns to existing datasets as new requirements surface. Schema changes involve only metadata changes and run fully concurrently with live workloads. Therefore, users can safely add columns to existing Cassandra databases while remaining confident that query performance will not degrade.