Version 3.1.0
Transactions

ACID properties

Transactions provide a powerful abstraction for multiple threads to operate on data concurrently because they have the following properties:

  • Atomicity: all or none of a transaction is completed.
  • Consistency: if each transaction maintains some property when considered separately, then the combined effect of executing the transactions concurrently will maintain the same property.
  • Isolation: developers can reason about transactions as if they run single-threaded.
  • Durability: once a transaction commits, its updates cannot be lost.

WiredTiger supports transactions with the following caveats to the ACID properties:

  • the maximum level of isolation supported is snapshot isolation. See Isolation levels for more details.
  • transactional updates are made durable by a combination of checkpoints and logging. See Checkpoint durability for information on checkpoint durability and Commit-level durability for information on commit-level durability.
  • each transaction's uncommitted changes must fit in memory: for efficiency, WiredTiger does not write to the log until a transaction commits.

Transactional API

In WiredTiger, transaction operations are methods off the WT_SESSION class.

Applications call WT_SESSION::begin_transaction to start a new transaction. Operations subsequently performed using that WT_SESSION handle, including operations on any cursors open in that WT_SESSION handle (whether opened before or after the WT_SESSION::begin_transaction call), are part of the transaction and their effects committed by calling WT_SESSION::commit_transaction, or discarded by calling WT_SESSION::rollback_transaction. Applications that use Application-specified Transaction Timestamps can utilize the WT_SESSION::prepare_transaction API as a basis for implementing a two phase commit protocol.

If WT_SESSION::commit_transaction returns an error for any reason, the transaction was rolled back, not committed.

When transactions are used, data operations can encounter a conflict and fail with the WT_ROLLBACK error. If this error occurs, transactions should be rolled back with WT_SESSION::rollback_transaction and the operation retried.

The WT_SESSION::rollback_transaction method implicitly resets all cursors in the session as if the WT_CURSOR::reset method was called, discarding any cursor position as well as any key and value.

/*
* Cursors may be opened before or after the transaction begins, and in
* either case, subsequent operations are included in the transaction.
* Opening cursors before the transaction begins allows applications to
* cache cursors and use them for multiple operations.
*/
error_check(session->open_cursor(
session, "table:mytable", NULL, NULL, &cursor));
error_check(session->begin_transaction(session, NULL));
cursor->set_key(cursor, "key");
cursor->set_value(cursor, "value");
switch (cursor->update(cursor)) {
case 0: /* Update success */
error_check(session->commit_transaction(session, NULL));
/*
* If commit_transaction succeeds, cursors remain positioned; if
* commit_transaction fails, the transaction was rolled-back and
* and all cursors are reset.
*/
break;
case WT_ROLLBACK: /* Update conflict */
default: /* Other error */
error_check(session->rollback_transaction(session, NULL));
/* The rollback_transaction call resets all cursors. */
break;
}
/*
* Cursors remain open and may be used for multiple transactions.
*/

Implicit transactions

If a cursor is used when no explicit transaction is active in a session, reads are performed at the isolation level of the session, set with the isolation key to WT_CONNECTION::open_session, and successful updates are automatically committed before the update operation returns.

Any operation consisting of multiple related updates should be enclosed in an explicit transaction to ensure the updates are applied atomically.

If an implicit transaction successfully commits, the cursors in the WT_SESSION remain positioned. If an implicit transaction fails, all cursors in the WT_SESSION are reset, as if WT_CURSOR::reset were called, discarding any position or key/value information they may have.

See Cursors and Transactions for more information.

Concurrency control

WiredTiger uses optimistic concurrency control algorithms. This avoids the bottleneck of a centralized lock manager and ensures transactional operations do not block: reads do not block writes, and vice versa.

Further, writes do not block writes, although concurrent transactions updating the same value will fail with WT_ROLLBACK. Some applications may benefit from application-level synchronization to avoid repeated attempts to rollback and update the same value.

Operations in transactions may also fail with the WT_ROLLBACK error if some resource cannot be allocated after repeated attempts. For example, if the cache is not large enough to hold the updates required to satisfy transactional readers, an operation may fail and return WT_ROLLBACK.

Isolation levels

WiredTiger supports read-uncommitted, read-committed and snapshot isolation levels; the default isolation level is read-committed.

  • read-uncommitted: Transactions can see changes made by other transactions before those transactions are committed. Dirty reads, non-repeatable reads and phantoms are possible.
  • read-committed: Transactions cannot see changes made by other transactions before those transactions are committed. Dirty reads are not possible; non-repeatable reads and phantoms are possible. Committed changes from concurrent transactions become visible when no cursor is positioned in the read-committed transaction.
  • snapshot: Transactions read the versions of records committed before the transaction started. Dirty reads and non-repeatable reads are not possible; phantoms are possible.

    Snapshot isolation is a strong guarantee, but not equivalent to a single-threaded execution of the transactions, known as serializable isolation. Concurrent transactions T1 and T2 running under snapshot isolation may both commit and produce a state that neither (T1 followed by T2) nor (T2 followed by T1) could have produced, if there is overlap between T1's reads and T2's writes, and between T1's writes and T2's reads.

The transaction isolation level can be configured on a per-transaction basis:

/* A single transaction configured for snapshot isolation. */
error_check(session->open_cursor(
session, "table:mytable", NULL, NULL, &cursor));
error_check(session->begin_transaction(session, "isolation=snapshot"));
cursor->set_key(cursor, "some-key");
cursor->set_value(cursor, "some-value");
error_check(cursor->update(cursor));
error_check(session->commit_transaction(session, NULL));

Additionally, the default transaction isolation can be configured and re-configured on a per-session basis:

/* Open a session configured for read-uncommitted isolation. */
error_check(conn->open_session(
conn, NULL, "isolation=read-uncommitted", &session));
/* Re-configure a session for snapshot isolation. */
error_check(session->reconfigure(session, "isolation=snapshot"));

Named Snapshots

Applications can create named snapshots by calling WT_SESSION::snapshot with a configuration that includes "name=foo". This configuration creates a new named snapshot, as if a snapshot isolation transaction were started at the time of the WT_SESSION::snapshot call.

Subsequent transactions can be started "as of" that snapshot by calling WT_SESSION::begin_transaction with a configuration that includes snapshot=foo. That transaction will run at snapshot isolation as if the transaction started at the time of the WT_SESSION::snapshot call that created the snapshot.

Named snapshots keep data pinned in cache as if a real transaction were running for the time that the named snapshot is active. The resources associated with named snapshots should be released by calling WT_SESSION::snapshot with a configuration that includes "drop=". See WT_SESSION::snapshot documentation for details of the semantics supported by the drop configuration.

Named snapshots are not durable: they do not survive WT_CONNECTION::close.

Application-specified Transaction Timestamps

Timestamp overview

Some applications have their own notion of time, including an expected commit order for transactions that may be inconsistent with the order assigned by WiredTiger. We assume that applications can represent their notion of a timestamp as an unsigned integral value of some size that generally increases over time. For example, a simple 64-bit integer could be incremented to generate transaction timestamps, if that is sufficient for the application.

The application's timestamp size is specified as a number of bytes at build time, with configure –with-timestamp-size=X. The default timestamp size is 8 bytes (i.e., 64 bits). Setting a size of zero disables transaction timestamp functionality. Timestamps are communicated to WiredTiger using a hexadecimal encoding, so the encoded value can be twice as long as the raw timestamp value.

Applications assign explicit commit timestamps to transactions, then read "as of" a timestamp. The timestamp mechanism operates in parallel with WiredTiger's internal transaction ID management. It is recommended that once timestamps are in use for a particular table, all subsequent updates also use timestamps.

Using transactions with timestamps

Applications that use timestamps will generally provide a timestamp at WT_SESSION::transaction_commit that will be assigned to all updates that are part of the transaction. WiredTiger also provides the ability to set a different commit timestamp for different sets of updates in a single transaction. This can be done by calling WT_SESSION::timestamp_transaction repeatedly to set a new commit timestamp between a set of updates for the current transaction. This gives the ability to commit updates with different read "as of" timestamps in a single transaction.

Setting a read timestamp in WT_SESSION::begin_transaction forces a transaction to run at snapshot isolation and ignore any commits with a newer timestamp.

Commit timestamps cannot be set in the past of any read timestamp that has been used. This is enforced by assertions in diagnostic builds, if applications violate this rule, data consistency can be violated.

The commits to a particular data item must be performed in timestamp order. If applications violate this rule, data consistency can be violated.

The WT_SESSION::prepare_transaction API is designed to be used in conjunction with timestamps and assigns a prepare timestamp to the transaction, which will be used for visibility checks until the transaction is committed or aborted. Once a transaction has been prepared the only other operations that can be completed are WT_SESSION::commit_transaction or WT_SESSION::rollback_transaction. The WT_SESSION::prepare_transaction API only guarantees that transactional conflicts will not cause the transaction to rollback - it does not guarantee that the transactions updates are durable. If a read operation encounters an update from a prepared transaction a WT_PREPARE_CONFLICT error will be returned indicating that it is not possible to choose a version of data to return until a prepared transaction is resolved, it is reasonable to retry such operations.

Managing global timestamp state

Applications that use timestamps need to manage some global state in order to allow WiredTiger to clean up old updates, and not make new updates durable until it is safe to do so. That state is managed using the WT_CONNECTION::set_timestamp API.

Setting an oldest timestamp in WT_CONNECTION::set_timestamp indicates that future read timestamps will be at least as recent as the oldest timestamp, so WiredTiger can discard history before the specified point. It is critical that the oldest timestamp update frequently or the cache can become full of updates, reducing performance.

Setting a stable timestamp in WT_CONNECTION::set_timestamp indicates a known stable location that is sufficient for durability. During a checkpoint the state of a table will be saved only as of the stable timestamp. Newer updates after that stable timestamp will not be included in the checkpoint. That can be overridden in the call to WT_SESSION::checkpoint. It is expected that the stable timestamp is updated frequently. Setting a stable location provides the ability, if needed, to rollback to this location by placing a call to WT_CONNECTION::rollback_to_stable. With the rollback, however, WiredTiger does not automatically reset the maximum commit timestamp it is tracking. The application should explicitly do so by setting a commit timestamp in WT_CONNECTION::set_timestamp.

Timestamp support in the extension API

The extension API, used by modules that extend WiredTiger via WT_CONNECTION::get_extension_api, is not timestamp-aware. In particular, WT_EXTENSION_API::transaction_oldest and WT_EXTENSION_API::transaction_visible do not take timestamps into account. Extensions relying on these functions may not work correctly with timestamps.