any-db-transaction

Transaction object for Any-DB adapters

any-db-transaction

A simple transaction helper for any-db compliant database adapters.

var anyDB = require('any-db')
var begin = require('any-db-transaction')
 
var connection = anyDB.createConnection(...)
 
// Callback-style 
begin(connection, function (errtransaction) {
  if (err) return console.error(err)
  // Do work using transaction 
  transaction.query(...)
  transaction.commit()
})
 
// Synchronous-style* 
var transaction = begin(connection)
transaction.on('error', console.error)
transaction.query(...)
transaction.commit()
 
// Or use a connection pool 
var pool = anyDB.createPool(...)
var transaction = begin(pool)
module.exports := begin(Queryable, statement: String?, Continuation<Transaction>?) => Transaction
 
Transaction := FSM & Queryable & {
  commit:   (Continuation?) => void
  rollback: (Continuation?) => void
}
module.exports := begin(Queryable, statement: String?, Continuation<Transaction>?) => Transaction

Transaction objects are are simple wrappers around a Connection that also implement the Queryable API, but guarantee that all queries take place within a single database transaction or not at all. Note that begin also understands how to acquire (and release) a connection from a ConnectionPool as well, so you can simply pass a pool to it: var tx = begin(pool)

By default, any queries that error during a transaction will cause an automatic rollback. If a query has no callback, the transaction will also handle (and re-emit) 'error' events for the Query instance. This enables handling errors for an entire transaction in a single place.

Transactions may also be nested by passing a Transaction to begin and these nested transactions can safely error and rollback without rolling back their parent transaction:

var parent = begin(connection)
var child = begin(parent)
child.query("some invalid sql")
child.on('error', function () {
  parent.query("select 1") // parent still works 
})

This feature relies on the SAVEPOINT support in your database. (In particular MySQL will doesn't have good support in very old versions). The use of savepoints also means there is no option to replace the statement used to begin the child transaction.

While the child transaction is in progress the parent transaction will queue any queries it receives until the child transaction either commits or rolls back, at which point it will process the queue. Be careful: it's quite possible to write code that deadlocks by waiting for a query in the parent transaction before committing the child transaction. For example:

// Do not do this! it will deadlock! 
 
var parent = begin(connection) // starts the transaction 
var child  = begin(parent)     // creates a savepoint 
 
parent.query('SELECT 1', function (err) {
  child.commit();
});

As stated previously, by default any queries that error during a transaction will cause an automatic rollback. This is to support the common pattern in which a transaction is a series of queries you either want to succeed or fail atomically.

There is another common pattern for transactions where you either create or update a record. Many databases support an INSERT OR REPLACE statement, but quite often you'd like an INSERT OR UPDATE construct instead.

Intuitively, a transaction can be used for this as well:

  1. Start a transaction
  2. Try an insert statement
  • If that succeeds, commit.
  • Otherwise, continue
  1. Try an update statement.
  • If that succeeds, commit.
  • Otherwise, roll back the transaction.

A transaction is unlikely to be the best choice here. The results of the first statement need to make it back to the client before it can decide whether to commit or try something else. Usually databases better support this kind of construct with nested queries, which avoid those roundtrips.

To facilitate this kind transaction use, automatic rollback of transactions can be disabled.

var tx = begin(conn, {autoRollback: false});
tx.query('Query that produces errors', function(err) {
    tx.query('another query');
});

Note: PostgreSQL does not allow you to use a transaction immediately after an error. However, you can get much the same behaviour by explicitly adding SAVEPOINT statements. A transaction with an error can be rolled back to a known good savepoint, and can be used from there onwards. You can achieve the same by using nested transactions.

var tx = begin(conn, {autoRollback: false});
var sp = begin(tx);
sp.query('query that might fail', function(err) {
  if (err) {
    tx.query('alternate queries');
  } else {
    sp.commit();
  }
});

Note that the failing query is performed on the "savepoint" child transaction, but the final query is perfomed on the outer/parent transaction.

Transactions are [FSM][] instances with 4 states: disconnected, connected, open, and closed:

[disconnected]
      ↓
 [connected]
   ↓  ↓  ↑
   ↓ [open]
   ↓   ↓
  [closed]

Every transaction starts out in the disconnected state, in which it will queue all tasks (queries, child transactions, commits and rollbacks) in the order they are received.

Once the transaction acquires a connection* it will transition to the connected state and begin processing it's internal task queue. While in this state any new tasks will still be added to the end of the queue. There are two possible transitions from the connected state:

  • connected → open - When all queued tasks have finished.
  • connected → closed - When a rollback or commit is encountered in the queue. This includes automatic rollbacks caused by query errors.

closed is a terminal state in which all further database operations result in errors. (The errors will either be sent to any callback provided or emitted as error events on the next tick).

In the open state, all database operations will be performed immediately. If a child transaction is started like var child = begin(parentTxn), the parent transaction will move back into the connected state (queueing any queries it receives) until the child completes, at which point it will resume processing it's own internal queue.

Transactions created from a Connection transition to connected before begin returns.

The Adapter instance used by the resource (connection or parent transaction) underlying this transaction.

(text: String, params: Array?, Continuation<Result>?) => Query

Maintains the same contract as Queryable.query but adds further guarantees that queries will be performed within the transaction or not at all. If the transaction has been committed or rolled back this method will fail by passing an error to the continuation (if provided) or emitting an 'error' event.

(Continuation<void>) => void

Issue a COMMIT (or RELEASE ... in the case of nested transactions) statement to the database. If a continuation is provided it will be called (possibly with an error) after the COMMIT statement completes. The transaction object itself will be unusable after calling commit().

(Continuation<void>) => void

The same as Transaction.commit but issues a ROLLBACK. Again, the transaction will be unusable after calling this method.

  • 'query', query - emitted immediately after .query is called on a connection via tx.query. The argument is a query object.

  • 'commit:start' - Emitted when .commit() is called.

  • 'commit:complete' - Emitted after the transaction has committed.

  • 'rollback:start' - Emitted when .rollback() is called.

  • 'rollback:complete' - Emitted after the transaction has rolled back.

  • 'close' - Emitted after rollback or commit completes.

  • 'error', err - Emitted under three conditions:

    1. There was an error acquiring a connection.
    2. Any query performed in this transaction emits an error that would otherwise go unhandled.
    3. Any of query, begin, commit, or rollback are called after the connection has already been committed or rolled back.

    Note that the 'error' event may be emitted multiple times! depending on the callback you are registering, you way want to wrap it using [once][].

A common pattern in web applications is start a transaction for each request and commit it before sending a response. Here is a simplified connect middleware that encapsulates this pattern:

module.exports = function unitOfWorkMiddleware (poolerrorHandler) {
  return function (reqresnext) {
    req.tx = pool.begin()
    // intercept writeHead to ensure we have completed our transaction before 
    // responding to the user 
    var writeHead = res.writeHead
    res.writeHead = function () {
       if (req.tx.state() != 'closed') {
         req.tx.commit(function (err) {
           if (err) {
             errorHandler(req, res, err)
           } else {
             writeHead.apply(res, arguments)
           }
         })
       } else {
         writeHead.apply(res, arguments)
       }
    }
    next()
  }
}

Here's an example where we stream all of our user ids, check them against an external abuse-monitoring service, and flag or delete users as necessary, if for any reason we only get part way through, the entire transaction is rolled back and nobody is flagged or deleted:

var pool = require('any-db').createPool(...)
 
// this is our external service 
var abuseService = require('./services').abuseService()
 
var tx = begin(pool)
tx.on('error', finished)
 
/*
Why query with the pool and not the transaction?
Because it allows the transaction queries to begin executing immediately,
rather than queueing them all up behind the initial SELECT.
*/
pool.query('SELECT id FROM users')
  .on('data', function (user) {
    if (tx.state() == 'closed') {
      // Do not make unneccessary requests 
      return
    }
    abuseService.checkUser(user.id, function (errresult) {
      if (err) return tx.handleError(err)
      // Errors from these queries will propagate up to the transaction object 
      if (result.flag) {
        tx.query('UPDATE users SET abuse_flag = 1 WHERE id = $1', [user.id])
      } else if (result.destroy) {
        tx.query('DELETE FROM users WHERE id = $1', [user.id])
      }
    })
  }).on('end', function () {
    tx.commit(finished)
  })
 
function finished (err) {
  if (err) console.error(err)
  else console.log('All done!')
}

License

2-clause BSD