Transactions

This section describes transactions as applied to Cypress, and static tables, locks, and versioning of Cypress objects.
The transactional model of dynamic tables is described in the Multiversioning and transaction processing of dynamic tables section.

General information

The YTsaurus system supports transactions, but differs in a number of ways from the classic transaction processing model:

Transactions can last minutes or hours.
Isolation is configurable.
There are no foreign table keys.
Transactions can affect other Cypress nodes in addition to tables.

Transactions are objects of the transaction type. A list of all system transactions can be found in Cypress at //sys/transactions.
A transaction can have another transaction as a parent. Transactions form a tree whose roots are transactions without parents, also called topmost transactions. A list of all topmost transactions is available at //sys/topmost_transactions.

YTsaurus ensures the following properties of the transaction processing system:

Atomicity. YTsaurus guarantees that the transaction will not be partially committed in the system. Either all or none of its sub-operations will be executed. Changing the data of a single Cypress node within a single command (for example, the set command) is atomic.
Consistency. Changes introduced by committing a transaction maintain data consistency in Cypress nodes and static tables.
Isolation. Unlike traditional transaction processing systems, YTsaurus enables you to choose isolation on a per-transaction basis. YTsaurus behavior corresponds to the Read committed isolation level when interacting with Cypress and Serializable when interacting with static tables within transactions.
Durability. YTsaurus guarantees the safety of changes after the transaction is committed. A single server hardware failure, power outage, or system shutdown cannot cause a loss of made changes. Transactions in the system can survive a system shutdown and continue running after the system is restored.

When you perform most actions in the system, you can specify which transaction those actions should be performed under. Even if no transaction is explicitly specified, the actions are atomic. YTsaurus starts a special system transaction and completes it at the end of the command execution.

Transactions are divided into master transactions and tablet transactions. Master transactions enable you to perform operations on master metainformation.
Tablet transactions enable you only to write data to dynamic tables. For more information about tablet transactions, see Dynamic tables.

Master transactions

Transactional processing within master servers applies to versionable objects. Examples of such objects are files, tables, and folders.

Creating a transaction

To create a transaction, run the start_tx command.
You can specify a parent transaction in the parent_id attribute and a transaction lifetime in the timeout attribute.
The upper transaction lifetime limit in the system is one hour. If you specify a timeout of more than one hour, it will be equal to the limit.
The reference point for transaction lifetime is the latter of start_tx moment or last ping_tx moment.

Extending a transaction lifetime

To extend a transaction lifetime, run the ping_tx command.
Each execution extends the transaction lifetime by a time interval equal to timeout. If the time since transaction creation or the last execution of the ping_tx command exceeds timeout, the transaction will be aborted.

Completing a transaction

You can abort a transaction using the abort_tx or successfully complete it using the commit_tx command.

If you abort a transaction, all nested transactions will also be stopped.
To successfully complete a transaction with nested transactions, run the commit_tx command on all nested transactions and then on the parent transaction.
In all other cases, completing a transaction cannot cause an error.

YTsaurus uses pessimistic locks, so possible conflicts are detected as they occur — when locks are taken and objects are created within transactions rather than when a transaction is completed.

Transaction attributes

In addition to the attributes inherent to all objects, transactions have the following attributes:

Attribute	Type	Description	Mandatory
`timeout`	`integer`	Transaction timeout in ms. May be omitted for some system transactions.	No
`title`	`string`	Text description string. This attribute is filled in automatically for all system transactions and for user transactions only if the user specifies it themselves when creating a transaction.	No
`last_ping_time`	`DateTime`	Time of the last transaction lifetime extension. May be missing for some system transactions.	No
`parent_id`	`Guid`	Parent transaction ID.	Yes
`start_time`	`DateTime`	Transaction creation time.	Yes
`nested_transaction_ids`	`array<Guid>`	A list of nested transaction IDs.	Yes
`staged_object_ids`	`array<Guid>`	A list of IDs of objects that the transaction temporarily owns.	Yes
`branched_node_ids`	`array<Guid>`	A list of branched Cypress node IDs.	Yes
`locked_node_ids`	`array<Guid>`	A list of locked Cypress node IDs.	Yes
`lock_ids`	`array<Guid>`	A list of IDs of locks created in the transaction.	Yes
`resource_usage`	`ResourceUsageMap`	An attribute that shows the use of resources in a given transaction for each affected account.	Yes

Note

Transactions created by the system always have the filled in title attribute. It contains a description of the operations associated with the transaction.
Users are also encouraged to use this attribute to describe the purpose of the transaction.

Locks

Versioning of Cypress nodes is related to the concept of locks. By taking a lock on the node, the transaction expresses its intention to work with the node according to the locking mode and, if successful, gets a guarantee that:

It is allowed to work with this node in the specified manner.
The node for this transaction is branched.

Note

The node may have been branched before, but if there is a lock, the branched version exists.
The opposite is not true: in certain scenarios, there may be branched but not locked nodes. An attempt to work with such a node will always result in taking a lock on it. Branching and taking a lock are related, but are different things.

A lock is a full-fledged object that has an ID. A list of all locks in the system is available at //sys/locks. We recommend using an address of the #lock-id form to access a specific lock.

A list of locks taken by a transaction is displayed in its lock_ids attribute.

Locking modes

The following locking modes are available: snapshot, exclusive, and shared.
A locking mode defines the list of allowed transaction actions, as well as the ability to take other locks:

snapshot: The transaction can read but not modify the node. The lock is used to take a read-only copy of the Cypress node state in the context of a transaction and freeze its state.

Note

A snapshot lock is taken only on the node itself, but not on the path to it in Cypress. If you continue accessing the node using its path, you can get a new node placed on the same path.
To guarantee that you can access the snapshot, use one of the following methods:
— Take a lock on the node id. In this case, the node can be changed between computing the id and taking the lock, so additional attempts may be required.
— Get the id of the snapshot node by lock_id. Use the lock method to get the lock_id, then read the snapshot node id from the #<lock_id>/@node_id path.

exclusive: The transaction can modify the node state. Other transactions cannot change the node.
shared: The transaction can modify a certain part of the node state. Other transactions can still change other parts of this node.

There are three standard scenarios for using this lock:
- Concurrently appending data to a table or file from multiple transactions. In this case, you can only append data, but not overwrite it.
- Concurrently creating several differently named subfolders within the same folder from multiple transactions. For example, transaction T1 can be started and create (or delete) the //tmp/a node and transaction T2 can be started and create (or delete) the //tmp/b node. Each of them will take a separate shared lock on //tmp. To detect conflicts, each lock has the child_key attribute indicating which key (subfolder) is locked by it.
- Concurrently creating several differently named attributes within the same node from multiple transactions. For example, transaction T1 can be started and set (change, delete) the //tmp/@a attribute and transaction T2 can be started and set (change, delete) the //tmp/@b attribute. To detect conflicts, each lock has the attribute_key attribute indicating which key (attribute name) is locked by it.

Note

Transactions can be nested. There can be more than one exclusive lock on the node due to nested transactions.

Implicit locks

A transaction can take locks either explicitly using the lock command or implicitly. Implicit taking of locks can occur in case of certain interactions with Cypress nodes, for example:

Creating a node is accompanied by taking an exclusive lock.
Writing to a table or file means taking a shared lock if writing to the end of the table and an exclusive lock if overwriting it.
Creating a new entry in the folder, as well as changing or deleting an existing entry is accompanied by taking a shared lock with the corresponding child_key.
Creating a new node attribute, as well as changing or deleting an existing attribute is accompanied by taking a shared lock with the corresponding attribute_key.

Lock compatibility

Some lock combinations can be taken concurrently. The formal rules are as follows:

A snapshot lock can always be taken. If this transaction has already taken a snapshot lock, an attempt to take the lock again is completed without errors and has no effect.
A shared or exclusive lock cannot be taken if the transaction or any of its ancestors has already taken a snapshot lock.
A shared or exclusive lock cannot be taken if another transaction that is not an ancestor of the given one has already taken an exclusive lock.
An exclusive lock cannot be taken if another transaction that is not an ancestor of the given one has already taken a shared lock.
A shared lock with the specified child_key cannot be taken if another transaction that is not an ancestor of the given one has already taken a shared lock with the same child_key.
A shared lock with the specified attribute_key cannot be taken if another transaction that is not an ancestor of the given one has already taken a shared lock with the same attribute_key.
A shared lock without child_key and attribute_key can be taken despite any other shared locks.

Lock operations

There are two commands to work with locks: lock and unlock.

The lock command enables you to take a lock on a Cypress node in a specified transaction.

The unlock command does the opposite: it removes all explicit locks from the node for a given transaction, both those already taken and those still in the lock queue.
The lock can only be removed if the locked branched version of the node contains no changes compared to the original. Consequently, an explicit snapshot lock can always be removed. Otherwise the command will end with an error.

Locks are automatically removed at the end of the transaction, whether successful or unsuccessful. Therefore, there is usually no need to remove them manually.
We recommend using the unlock command only when you need to take and release locks without completing the transaction. Such transactions are usually designed to manage the synchronization of third-party services.

Lock queue

Each Cypress node may have a lock queue.

By default, the lock command tries to take a lock and returns an error if the node is already locked. If you specify the waitable parameter equal to true in the lock command, the lock will be queued.
To find out whether a lock is in the queue, request the state attribute. If the lock is in the queue, it will be pending, if not, it will be acquired.

Attention!

For a lock to be actually taken, for example exclusively, its state must become acquired.
To take a lock:

Call the lock command.
Monitor the state in the cycle until it takes on the acquired value.

Lock attributes

In addition to the attributes inherent to all objects, locks have the following attributes:

Attribute	Type	Description
`state`	`string`	Lock state: `pending` or `acquired`.
`transaction_id`	`Guid`	ID of the transaction that took the lock.
`mode`	`string`	Locking mode: `shared`, `exclusive`, or `snapshot`.
`child_key`	`string`	The key on which the lock is taken. For the `shared` type only.
`attribute_key`	`string`	The name of the attribute on which the lock is taken (for the `shared` type only)

Versioning

Changing the node state by a transaction is the following three-phase process:

Transaction T takes the lock on node N. Version N:T appears for node N and it is formed as follows:
- For a snapshot lock, it branches off from version N:T' where T' is the closest ancestor of T that branched off N.
- For shared and exclusive locks, N:T'' branches off from N:T' where T'' is the child of T' ; N:T''' branches off from N:T'' where T''' is the child of T'' and so on up to N:T. In other words, a chain of branched node versions is created for each transaction from T' to T.
If there is no such T', the real version of N is used. It does not matter for this description whether the lock was taken explicitly or implicitly.
Transaction T works with node N, with the actual changes accumulated in its version N:T.
Transaction T completes successfully and the changes it made to N:T, if any, get merged into the version from which version N:T was branched. Thus, these changes become visible to the parent transaction or to all if transaction T was a topmost transaction.

Using transactions in operations

When starting operations, the scheduler creates a set of transactions to provide some atomicity of data processing in the operation. Learn more about the work schema in the Transaction processing of data section.