Because CockroachDB is a distributed transactional database, the path queries take is dramatically different from many other database architectures. To help familiarize you with CockroachDB's internals, this guide covers what that path actually is.
If you've already read the CockroachDB architecture documentation, this guide serves as another way to conceptualize how the database works. This time, instead of focusing on the layers of CockroachDB's architecture, we're going to focus on the linear path that a query takes through the system (and then back out again).
To get the most out of this guide, we recommend beginning with the architecture documentation's overview and progressing through all of the following sections. This guide provides brief descriptions of each component's function and links to other documentation where appropriate, but assumes the reader has a basic understanding of the architecture in the first place.
This guide is organized by the physical actors in the system, and then broken down into the components of each actor in the sequence in which they're involved.
Here's a brief overview of the physical actors, in the sequence with which they're involved in executing a query:
- SQL Client sends a query to your cluster.
- Load Balancing routes the request to CockroachDB nodes in your cluster, which will act as a gateway.
- Gateway is a CockroachDB node that processes the SQL request and responds to the client.
- Leaseholder is a CockroachDB node responsible for serving reads and coordinating writes of a specific range of keys in your query.
- Raft leader is a CockroachDB node responsible for maintaining consensus among your CockroachDB replicas.
Once the transaction completes, queries traverse these actors in approximately reverse order. We say "approximately" because there might be many leaseholders and Raft leaders involved in a single query, and there is little-to-no interaction with the load balancer during the response.
SQL Client/Postgres Wire Protocol
To begin, a SQL client (e.g., an app) performs some kind of business logic against your CockroachDB cluster, such as inserting a new customer record.
This request is sent over a connection to your CockroachDB cluster that's established using a PostgreSQL driver.
Load Balancing & Routing
Modern architectures require distributing your cluster across machines to improve throughput, latency, and uptime. This means queries are routed through load balancers, which choose the best CockroachDB node to connect to. Because all CockroachDB nodes have perfectly symmetrical access to data, this means your load balancer can connect your client to any node in the cluster and access any data while still guaranteeing strong consistency.
Your architecture might also have additional layers of routing to enforce regulatory compliance, such as ensuring GDPR compliance.
Once your router and load balancer determine the best node to connect to, your client's connection is established to the gateway node.
The gateway node handles the connection with the client, both receiving and responding to the request.
SQL parsing & planning
Given that CockroachDB is a distributed database, though, it's also important to take a cluster's topology into account, so the logical plan is then converted into a physical plan—this means sometimes pushing operations onto the physical machines that contain the data.
While CockroachDB presents a SQL interface to clients, the actual database is built on top of a key-value store. To mediate this, the physical plan generated at the end of SQL parsing is passed to the SQL executor, which executes the plan by performing key-value operations through the
TxnCoordSender. For example, the SQL executor converts
TxnCoordSender provides an API to perform key-value operations on your database.
On its back end, the
TxnCoordSender performs a large amount of the accounting and tracking for a transaction, including:
- Accounts for all keys involved in a transaction. This is used, among other ways, to manage the transaction's state.
- Packages all key-value operations into a
BatchRequest, which are forwarded on to the node's
The gateway node's
BatchRequests from the
TxnCoordSender. It dismantles the initial
BatchRequest by taking each operation and finding which physical machine should receive the request for the range—known as the range's leaseholder. The address of the range's current leaseholder is readily available in both local caches, as well as in the cluster's
BatchRequests are reassembled into new
BatchRequests containing the address of the range's leaseholder.
All write operations also propagate the leaseholder's address back to the
TxnCoordSender, so it can track and clean up write operations as necessary.
DistSender sends out the first
BatchRequest for each range in parallel. As soon as it receives a provisional acknowledgment from the leaseholder node’s evaluator (details below), it sends out the next
BatchRequest for that range.
DistSender then waits to receive acknowledgments for all of its write operations, as well as values for all of its read operations. However, this wait isn't necessarily blocking, and the
DistSender may still perform operations with ongoing transactions.
The gateway node's
DistSender tries to send its
BatchRequests to the replica identified as the range's leaseholder, which is a single replica that serves all reads for a range, as well as coordinates all writes. Leaseholders play a crucial role in CockroachDB's architecture, so it's a good topic to make sure you're familiar with.
Because the leaseholder replica can shift between nodes, all nodes must be able to return a request for any key, returning a response indicating one of these scenarios:
No Longer Leaseholder
If a node is no longer the leaseholder, but still contains a replica of the range, it denies the request but includes the last known address for the leaseholder of that range.
Upon receipt of this response, the
DistSender will update the header of the
BatchRequest with the new address, and then resend the
BatchRequest to the newly identified leaseholder.
No Longer Has/Never Had Range
If a node doesn't have a replica for the requested range, it denies the request without providing any further information.
In this case, the
DistSender must look up the current leaseholder using the cluster's
Once the node that contains the leaseholder of the range receives the
BatchRequest, it begins processing it, and progresses onto checking the timestamp cache.
The timestamp cache tracks the highest timestamp (i.e., most recent) for any read operation that a given range has served.
Each write operation in a
BatchRequest checks its own timestamp versus the timestamp cache to ensure that the write operation has a higher timestamp; this guarantees that history is never rewritten and you can trust that reads always served the most recent data. It's one of the crucial mechanisms CockroachDB uses to ensure serializability. If a write operation fails this check, it must be restarted at a timestamp higher than the timestamp cache's value.
Operations in the
BatchRequest are serialized through the leaseholder's latch manager.
This works by giving each write operation a latch on a row. Any reads or writes that come in after the latch has been granted on the row must wait for the write to complete, at which point the latch is released and the subsequent operations can continue.
The batch evaluator ensures that write operations are valid. Our architecture makes this fairly trivial. First, the evaluator can simply check the leaseholder's data to ensure the write is valid; because it has coordinated all writes to the range, it must have the most up-to-date versions of the range's data. Secondly, because of the latch manager, each write operation is guaranteed to uncontested access to the range (i.e., there is no contention with other write operations).
If the write operation is valid according to the evaluator, the leaseholder sends a provisional acknowledgment to the gateway node's
DistSender; this lets the
DistSender begin to send its subsequent
BatchRequests for this range.
Importantly, this feature is entirely built for transactional optimization (known as transaction pipelining). There are no issues if an operation passes the evaluator but doesn't end up committing.
Writing the transaction record
The very first range that a transaction interacts with also receives a write for a key known as its transaction record. This includes a UUID to uniquely identify the transaction, as well as the canonical record of its state—e.g.,
COMMITTED. This record is used by all operations to determine the state of the transaction. For example, when a write is blocked by another write, it periodically checks its own transaction record to ensure that it hasn't been aborted.
As soon as the write for the transaction record completes (i.e., achieves consensus through Raft), it begins communicating with the gateway node's
TxnCoordSender, which regularly heartbeats the transaction to keep it alive (i.e., in a
PENDING state). If the heartbeating from the
TxnCoordSender stops, the
TxnRecord moves itself into
This write is handled like all other writes, which we'll discuss in more detail in a subsequent section.
Reads from RocksDB
All operations (including writes) begin by reading from the local instance of RocksDB to check for write intents for the operation's key. We talk much more about write intents in the transaction layer of the CockroachDB architecture, which is worth reading, but a simplified explanation is that these are provisional, uncommitted writes that express that some other concurrent transaction plans to write a value to the key.
What we detail below is a simplified version of the CockroachDB transaction model. For more detail, check out the transaction architecture documentation.
Resolving Write Intents
If an operation encounters a write intent for a key, it attempts to "resolve" the write intent by checking the state of the write intent's transaction. If the write's intent's transaction is...
COMMITTED, this operation converts the write intent to a regular key-value pair, and then proceeds as if it had read that value instead of a write intent.
ABORTED, this operation discards the write intent and reads the next-most-recent value from RocksDB.
PENDING, the new transaction attempts to "push" the write intent's transaction by moving that transaction's timestamp forward (i.e., ahead of this transaction's timestamp); however, this only succeeds if the write intent's transaction has become inactive.
If the push succeeds, the operation continues.
If this push fails (which is the majority of the time), this transaction goes into the
TxnWaitQueue on this node. The incoming transaction can only continue once the blocking transaction completes (i.e., commits or aborts).
Check out our architecture documentation for more information about CockroachDB's transactional model.
If the read doesn't encounter a write intent and the key-value operation is meant to serve a read, it can simply use the value it read from the leaseholder's instance of RocksDB. This works because the leaseholder had to be part of the Raft consensus group for any writes to complete, meaning it must have the most up-to-date version of the range's data.
The leaseholder aggregates all read responses into a
BatchResponse that will get returned to the gateway node's
As we mentioned before, each read operation also updates the timestamp cache.
After guaranteeing that there are no existing write intents for the keys,
BatchRequest's key-value operations are converted to Raft operations and have their values converted into write intents.
The leaseholder then proposes these Raft operations to the Raft group leader. The leaseholder and the Raft leader are almost always the same node, but there are situations where the roles might drift to different nodes. However, when the two roles are not collocated on the same physical machine, CockroachDB will attempt to relocate them on the same node at the next opportunity.
CockroachDB leverages Raft as its consensus protocol. If you aren't familiar with it, we recommend checking out the details about how CockroachDB leverages Raft, as well as learning more about how the protocol works at large.
In terms of executing transactions, the Raft leader receives proposed Raft commands from the leaseholder. Each Raft command is a write that is used to represent an atomic state change of the underlying key-value pairs stored in RocksDB.
For each command the Raft leader receives, it proposes a vote to the other members of the Raft group.
Once the command achieves consensus (i.e., a majority of nodes including itself acknowledge the Raft command), it is committed to the Raft leader’s Raft log and written to RocksDB. At the same time, the Raft leader also sends a command to all other nodes to include the command in their Raft logs.
Once the leader commits the Raft log entry, it’s considered committed. At this point the value is considered written, and if another operation comes in and performs a read on RocksDB for this key, they’ll encounter this value.
Note that this write operation creates a write intent; these writes will not be fully committed until the gateway node sets the transaction record's status to
On the way back up
Now that we have followed an operation all the way down from the SQL client to RocksDB, we can pretty quickly cover what happens on the way back up (i.e., when generating a response to the client).
- Once the leaseholder applies a write to its Raft log, it sends an commit acknowledgment to the gateway node's
DistSender, which was waiting for this signal (having already received the provisional acknowledgment from the leaseholder's evaluator).
- The gateway node's
DistSenderaggregates commit acknowledgments from all of the write operations in the
BatchRequest, as well as any values from read operations that should be returned to the client.
- Once all operations have successfully completed, the
DistSenderchecks the transaction record. There are a few situations that might arise here:
- If the transaction's timestamp was moved, the transaction performs a read refresh to see if any values it needed have been changed. If the read refresh is successful, the transaction can commit at the pushed timestamp. If the read refresh fails, the transaction must be restarted.
- If the transaction is in an
DistSendersends a response indicating as much, which ends up back at the SQL interface.
- If the transaction is still
PENDING, it's moved to
COMMITTED. Note that this modification of the transaction record is handled as an update to the key-value pair, meaning it must go through Raft like all other write operations. Once the transaction record is moved to
COMMITTED, the transaction is considered committed.
DistSenderpropagates any values that should be returned to the client (e.g., reads or the number of affected rows) to the
TxnCoordSender, which in turn responds to the SQL interface with the value. The
TxnCoordSenderalso begins asynchronous intent cleanup by sending a request to the
DistSenderto convert all write intents it created for the transaction to fully committed values. However, this process is largely an optimization; if any operation encounters a write intent, it checks the write intent's transaction record. If the transaction record is
COMMITTED, the operation can perform the same cleanup and convert the write intent to a fully committed value.
- The SQL interface then responds to the client, and is now prepared to continue accepting new connections.