Two-Phase Commit

2PC is an experimental feature and is likely not robust enough to be considered production-ready.
Transaction commit is much slower when using 2PC. The authors of Vitess recommend that you design your VSchema so that cross-shard updates (and 2PC) are not required.

Vitess 2PC allows you to perform atomic distributed commits. The feature is implemented using traditional MySQL transactions, and hence inherits the same guarantees. With this addition, Vitess can be configured to support the following three levels of atomicity:

  1. Single database: At this level, only single database transactions are allowed. Any transaction that tries to go beyond a single database will fail.
  2. Multi database: A transaction can span multiple databases, but the commit will be best effort. Partial commits are possible.
  3. 2PC: This is the same as Multi-database, but the commit will be atomic.

2PC commits are more expensive than multi-database because the system has to save away the statements before starting the commit process, and also clean them up after a successful commit. This is the reason why it is a separate option instead of being always on.

Isolation #

2PC transactions guarantee atomicity: either the whole transaction commits, or it is rolled back entirely. It does not guarantee Isolation (in the ACID sense). This means that a third party that performs cross-database reads can observe partial commits while a 2PC transaction is in progress.

Guaranteeing ACID Isolation is very contentious and has high costs. Providing it by default would have made Vitess impractical for the most common use cases.

Configuring VTGate #

The atomicity policy is controlled by the transaction_mode flag. The default value is multi and will set all transactions to multi-database mode.

To enforce single-database transactions, the VTGates can be started by specifying transaction_mode=single.

To enable 2PC, the VTGates need to be started with transaction_mode=twopc. The VTTablets will require additional flags, which will be explained below.

The VTGate transaction_mode flag decides what to allow by default. The application can then override that global default for an individual transaction using SET transaction_mode="<mode>"; when necessary or appropriate.

Driver APIs #

The way to request atomicity from the application is driver-specific.

MySQL Protocol #

Clients can set the transaction mode via a session-variable:

set transaction_mode='twopc';

gRPC Clients #

Go driver #

For the Go driver, you request the atomicity by adding it to the context using the WithAtomicity function. For more details, please refer to the respective GoDocs.

Python driver #

For Python, the begin function of the cursor has an optional single_db flag. If the flag is True, then the request is for a single-db transaction. If False (or unspecified), then the following commit call’s twopc flag decides if the commit is 2PC or Best Effort (multi).

Adding support in a new driver #

The VTGate RPC API extends the Begin and Commit functions to specify atomicity. The API mimics the Python driver: The BeginRequest message provides a single_db flag and the CommitRequest message provides an atomic flag which is synonymous to twopc.

Configuring VTTablet #

The following flags need to be set to enable 2PC support in VTTablet:

  • twopc_enable: This flag needs to be turned on.
  • twopc_coordinator_address: This should specify the address (or VIP) of the VTGate that VTTablet will use to resolve abandoned transactions.
  • twopc_abandon_age: This is the time in seconds that specifies how long to wait before asking a VTGate to resolve an abandoned transaction.

With the above flags specified, every primary VTTablet also turns into a watchdog. If any 2PC transaction is left lingering for longer than twopc_abandon_age seconds, then VTTablet invokes VTGate and requests it to resolve it. Typically, the abandon_age needs to be substantially longer than the time it takes for a typical 2PC commit to complete (10s of seconds).

Configuring MySQL #

The usual default values of MySQL are sufficient. However, it is important to verify that the wait_timeout (28800) has not been changed. If this value was changed to be too short, then MySQL could prematurely kill a prepared transaction causing data loss.

Monitoring #

A few additional variables have been added to /debug/vars. Failures described below should be rare. But these variables are present so you can build an alert mechanism if anything were to go wrong.

Critical failures #

The following errors are not expected to happen. If they do, it means that 2PC transactions have failed to commit atomically:

  • InternalErrors.TwopcCommit: This is a counter that shows the number of times a prepared transaction failed to fulfil a commit request.
  • InternalErrors.TwopcResurrection: This counter is incremented if a new primary failed to resurrect a previously prepared (and unresolved) transaction.

Alertable failures #

The following failures are not urgent, but require investigation:

  • InternalErrors.WatchdogFail: This counter is incremented if there are failures in the watchdog thread of VTTablet. This means that the watchdog is not able to alert VTGate of abandoned transactions.
  • Unresolved.Prepares: This is a gauge that is set based on the number of lingering Prepared transactions that have been alive for longer than 5x the abandon age. This usually means that a distributed transaction has repeatedly failed to resolve. A more serious condition is when the metadata for a distributed transaction has been lost and this Prepare is now permanently orphaned.

Repairs #

If any of the alerts fire, it is time to investigate. Once you identify the dtid or the VTTablet that originated the alert, you can navigate to the /twopcz URL. This will display three lists:

  • Failed Transactions: A transaction reaches this state if it failed to commit. The only action allowed for such transactions is that you can discard it. However, you can record the DMLs that were involved and have someone come up with a plan to repair the partial commit.
  • Prepared Transactions: Prepared transactions can be rolled back or committed. Prepared transactions must be remedied only if their root Distributed Transaction has been lost or resolved.
  • Distributed Transactions: Distributed transactions can only be Concluded (marked as resolved).