Skip to content
This repository has been archived by the owner on Jan 10, 2023. It is now read-only.

Latest commit

 

History

History
191 lines (152 loc) · 7.08 KB

05_transaction.md

File metadata and controls

191 lines (152 loc) · 7.08 KB

Lovefield Specification

5. Transaction

5.1 Execution Plan

Queries are executed by the execution engine when either lf.query.Builder#exec, lf.Transaction#exec, or lf.Transaction#attach is invoked. The queries will be converted into execution plans understandable by the query engine. Lovefield offers explain() which will generate an execution plan but will not execute it. This is very useful for debugging a slow query.

var p = db.getSchema().table('Photo');
var a = db.getSchema().table('Album');
var query = db.select().
    from(p).
    innerJoin(a, p.albumId.eq(a.id)).
    where(a.id.eq('1'));
console.log(query.explain());  // DEBUG only

5.2 Implicit Transaction

Most of the time, queries are executed in implicit transactions due to the concise syntax:

ds.connect().then(function(db) {
  var item = db.getSchema().table('Item');

  // SELECT * FROM Item;
  return db.select().from(item).exec();
});

In this simple example, a select query is built via a query builder returned by db.select(). The query builder uses various augmentation functions to complete the query context without risk of SQL injection, and at the end the exec() method is called and the query is dispatched to query engine. Query engine will create an implicit transaction for this query, execute it, and return the results.

5.3 Explicit Transactions

Explicit transactions in Lovefield are read-write transactions. They are designed to atomically execute multiple queries. The simple usage of an explicit transaction is demonstrated in the following:

// Get a transaction object first.
var tx1 = db.createTransaction();

// exec in order: query 1 first then query2, guaranteed snapshot
tx1.exec([query1, query2]);

// exec in order: query 3 first then query4
query3.exec().then(function() {
  query4.exec();
});

// exec in parallel (syntactically): tx1, query3, query5, query6
query5.exec();
query6.exec();

var tx2 = db.createTransaction();
tx2.exec([query7, query8]);

The exec() function of a transaction also returns a promise, just like other queries. The difference is that the transaction guarantees all queries flush to persistent store atomically. If a transaction is created for select() queries, the select() queries will be carried out from the same snapshot.

Transactions are always run in the order of the exec() call received by the library. So the following situation can happen just like their SQL counterpart:

t0 ---------- t1 ---------- t2 ---------- t3 ----------->
S0                          S1
|create tx1
              |create tx2
                            |tx2 exec and committed
                                          |tx1 exec and committed

Although transaction tx1 is created before tx2, because the calling of its exec() is after tx2 been committed, tx1 will operate on the snapshot S1 instead of the snapshot S0 of its creation. In short, please treat exec() of a transaction object as both BEGIN TRANSACTION and COMMIT.

When a transaction (either implicit or explicit) is resolved, no further queries can be made within that transaction. Trying to reuse the transaction object will result in error.  This implies any conditional behavior must either span multiple transactions or be composed out of the available primitives within a transaction.

In many cases, queries in the same transaction must be executed in sequence, and the following queries are referencing the results from previous queries. In Lovefield, this is done through transaction attachment, as demonstrated below:

var schema = db.getSchema();
var e = schema.table('Employee');
var v = schema.table('Vacations');

// Get a transaction object as usual.
var tx = db.createTransaction();

// Secure the scope of queries so that there will be no surprise.
// All tables will be exclusively locked. See Concurrency Control section.
tx.begin([e, v]).then(function() {
  var q1 = db.select(e.id).from(e).where(e.hireDate.gt(someDate));

  // Attach will actually run the query in memory and get back the results.
  return tx.attach(q1);
}).then(function(results) {
  var ids = results.map(function(row) {
    return row['id'];
  });

  var q2 = db.update(v).set(v.days, 15).where(v.empId.in(ids));
  return tx.attach(q2);
}).then(function() {
  // Commit the transaction, which writes everything into database.
  // Remember, commit() is an asynchronous call that returns a Promise.
  return tx.commit();
}).then(function() {
  // ...
});

var tx2 = db.createTransaction();
tx2.begin([e]).then(function() {
  return tx.attach(
      db.update(e).set(e.location, 'MTV').where(e.id.lt(1000)));
}).then(function() {
  // exec() can be used to commit the transaction, too.
  return tx.exec(
      [db.update(e).set(e.location, 'LAX').where(e.id.gte(1000))]);
});

var tx3 = db.createTransaction();
tx3.begin([v]).then(function() {
  return tx.attach(db.update(v).set(v.days, 0);
}).then(function() {
  // Canceling everyone's vacation is not really a good idea.
  // Remember, rollback() is an asynchronous call that returns a Promise.
  return tx.rollback();
});

The exec(), commit(), and rollback() call will make the transaction be in the termination state, which means that all member functions of the transaction object will throw error if called after termination.

The transactions are defined by interface lf.Transaction.

5.4 Concurrency Control

Lovefield offers table-level locking. There are three types of locks:

  • Shared: a reader lock that can be granted to multiple readers.
  • Reserved: a try-writer lock. This prevents granting new Shared or Reserved lock for the target table, but the target table is not modified yet.
  • Exclusive: a writer lock. This prevents granting any new lock against the target table since the table is going to be modified. The table can only be modified when an Exclusive lock is acquired.

These locks are created in the scenarios listed below:

Functions causing lock creation Lock created
exec() of a select() query Shared
exec() of an insert() query Reserved
exec() of an insertOrReplace() query Reserved
exec() of an update() query Reserved
exec() of an delete() query Reserved
begin() or exec() of a transaction Reserved

All reserved locks will be promoted to Exclusive locks automatically by the query runner. The exclusive locks promoted from locks created by exec() method will be automatically released once Lovefield has completed necessary data writing. Locks created by transaction's begin() method will not be promoted nor released until rollback() or commit() are called on that transaction. This implies the possibility of deadlock when multiple closures are attempting to write the database via transactions. The users are responsible for preventing and detecting these user code generated deadlocks.