Transaction Processing in PostgreSQL

Transaction Processing in PostgreSQL

Tom Lane
Great Bridge, LLC
[email protected]
1
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Outline

Introduction
• What is a transaction?
User’s view
• Multi-version concurrency control
Implementation
• Tuple visibility
• Storage management
• Locks

2
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

PostgreSQL system overview

Client Processes Server Processes

Postmaster
Client Daemon
Application Process
Initial
DB Requests Connection
Spawn
and Results Request
Server
via and
Process
Library API Authentication

Client Postgres
Interface SQL Queries
Server
Library and Results (Backend)

3
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

PostgreSQL system overview

Server Processes Shared Memory Unix System

Postmaster Create
Daemon Shared Kernel
Process Disk Disk
Buffers Buffers
Spawn
Server
Process
Shared
Read/ Disk
Postgres Write Tables
Server Storage
(Backend)

• Database files are accessed through shared buffer pool

• Hence, two backends can never see inconsistent views of a file
• Unix kernel usually provides additional buffering
4
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

What is a transaction, anyway?

Definition: a transaction is a group of SQLcommands whose results will be made

visible to the rest of the system as aunit when the transaction commits --- or not
at all, if the transaction aborts.

Transactions are expected to be atomic, consistent,isolated, and durable.

• Postgres does not support distributed transactions, so all commandsof a transaction

are executed by one backend.
• We don’t currently handle nested transactions, either.

5
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

The ACID test: atomic, consistent, isolated, durable

Atomic: results of a transaction are seen entirely or not at all within other transactions.
(A transaction need not appear atomic to itself.)

Consistent: system-defined consistency constraints areenforced on the results of

transactions. (Not going to discuss constraintchecking today.)

Isolated: transactions are not affected by the behavior ofconcurrently-running

transactions.
Stronger variant: serializable. If the final resultsof a set of concurrent transactions
are the same as if we’d run thetransactions serially in some order (not necessarily
any predetermined order), then we say the behavior is serializable.

Durable: once a transaction commits, its results will not belost regardless of
subsequent failures.
6
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

But how can thousands of changes be made "atomically"?

• The actual tuple insertions/deletions/updates are all markedas done by transaction N

as they are being made. Concurrently running backends ignore the changes
because they know transaction N is not committed yet. When the transactioncommits,
all those changes become logically visible at once.
• The control file pg_log contains 2 status bits pertransaction ID, with possible states
in progress, committed,aborted. Setting those two bits to thevalue committed is
the atomic action that marks a transaction committed.
• An aborting transaction will normally set its pg_log statusto aborted. But even if
the process crashes without havingdone so, everything is safe. The next time some
backend checks the state ofthat transaction, it will observe that the transaction is
marked inprogress but is not running on any backend, deduce that it crashed,and
update the pg_log entry to aborted on its behalf.No changes are needed
in any table data file during abort.

7
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

But is it really atomic and durable, even if the system crashes?

Well ... that depends on how much you trust your kernel and hard disk.

• Postgres transactions are only guaranteed atomic if a disk page write isan atomic
action. On most modern hard drives that’s true if a page is aphysical sector, but most
people run with disk pages configured as 8K or so,which makes it a little more dubious
whether a page write is all-or-nothing.

• pg_log is safe anyway since we’re only flipping bits init, and both bits of a
transaction’s status must be in the same sector.

• But when moving tuples around in a data page, there’s a potential fordata corruption
if a power failure should manage to abort the page writepartway through (perhaps only
some of the component sectors get written).This is one reason to keep page sizes
small ... and to buy a UPS for your server!
8
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Working through the Unix kernel costs us something, too

It’s critical that we force a transaction’s data page changes down to diskbefore we write
pg_log. If the disk writes occur in the wrongorder, a power failure could leave us with a
transaction that’smarked committed in pg_log but not all ofwhose data changes are
reflected on disk --- thus failing the atomicity test.

• Unix kernels allow us to force the correct write order via fsync(2), butthe performance
penalty of fsync’ing many files is pretty high.

• We’re looking at ways to avoid needing so many fsync()s, but that’s adifferent talk.

9
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

User’s view: multi-version concurrency control

A PostgreSQL application sees the following behavior of concurrenttransactions:

• Each transaction sees a snapshot (database version) as of its starttime,

no matter what other transactions are doing while it runs

• Readers do not block writers, writers do not block readers

• Writers only block each other when updating the same row

10
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Concurrent updates are tricky

Consider this example: transaction A does

UPDATE foo SET x = x + 1 WHERE rowid = 42

and before it commits,transaction B comes along and wants to do the same thing
on the same row.

• B clearlymust wait to see if A commits or not.

• If A aborts then B can go ahead,using the pre-existing value of x.
• But if A commits, what then?
• Usingthe old value of x will yield a clearly unacceptableresult: x ends up
incremented by 1 not 2 after both transactions commit.
• But if B is allowed to increment the new valueof x, then B is reading data committed
since it began execution. This violates the basic principle oftransaction isolation.

11
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Read committed vs. serializable transaction level

PostgreSQL offers two answers to the concurrent-update problem(out of four

transaction isolation levels defined in the ISO SQL standard):
Read committed level: allow B to use new tuple as inputvalues (after checking
to ensure new tuple still satisfies query’s WHERE clause). Thus, B isallowed to
see just this tuple of A’s results.
Serializable level: abort B with "not serializable" error.Client application must redo
the whole transaction B, which will then be allowedto see the new value of x under
strict serializable-behavior rules.

• Serializable level is logically cleaner but requires more code inapplication, so by

default we run in read-committed level which usually produces the desiredbehavior.
• In either case a pure SELECT transaction only sees data committed beforeit started.
It’s just updates and deletes that are interesting.

12
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

How it’s implemented

"O say, can you see that tuple?"

The most fundamental implementation concept is tuplevisibility: which versions

of which table rows are seen by which transactions.

Ignoring tuples you’re not supposed to be able to see is the key tomaking
transactions appear atomic.

Definition: a tuple is a specific stored object ina table,representing one version

of some logical table row. A row may exist inmultiple versions simultaneously.

13
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Non-overwriting storage management

We must store multiple versions of every row. A tuple can be removed onlyafter
it’s been committed as deleted for long enough that no activetransaction
can see it anymore.

Fortunately, PostgreSQL has always practiced "non overwriting" storage

management: updated tuples are appended to the table, and older versionsare
removed sometime later.

Currently, removal of long-dead tuples is handled bya VACUUM maintenance

command that must be issuedperiodically. We are looking at ways to reduce
need for VACUUM by recycling dead tuples on-the-fly.

14
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Per-tuple status information

Tuple headers contain:

• xmin: transaction ID of inserting transaction
• xmax: transaction ID of replacing/deleting transaction (initially NULL)
• forward link: link to newer version of same logical row, if any
Basic idea: tuple is visible if xmin is valid and xmax is not. "Valid"means
"either committed or the current transaction".

If we plan to update rather than delete, we first add new version of rowto table,
then set xmax and forward link in old tuple. Forward link willbe needed by
concurrent updaters (but not by readers).
To avoid repeated consultation of pg_log, there are alsosome statusbits that indicate
"known committed" or "known aborted" for xmin and xmax.These are set by the first
backend that inspects xmin or xmax after thereferenced transaction commits/aborts.

15
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

"Snapshots" filter away active transactions

If Transaction A commits while Transaction B is running, we don’t want Bto

suddenly start seeing A’s updates partway through.

• Hence, we make a listat transaction start of which transactions are currently being
run by other backends.(Cheap shared-memory communication is essential here: we
just look in ashared-memory table, in which each backend records its current
transactionnumber.)
• These transaction IDs will never be considered validby the current transaction,
even if they are shown to be committed in pg_log or on-row status bits.
• Nor will a transaction with ID higher than the current transaction’sever be
considered valid.
• These rules ensure that no transaction committing after the currenttransaction’s
start will be considered committed.
• Validity is in the eye of the beholder.
16
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Table-level locks: still gotta have ’em for some things

Even though readers and writers don’t block each other under MVCC, we stillneed
table-level locking.

This exists mainly to prevent the entire table frombeing altered or deleted
out from under readers or writers.

We also offer various lock levels for application use (mainly forporting applications
that take a traditional lock-based approach toconcurrency).

17
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Types of locks

Lock type Acquired by system for Conflicts with

1 AccessShareLock SELECT 7

2 RowShareLock SELECT FOR UPDATE 6,7

3 RowExclusiveLock UPDATE, INSERT, DELETE 4,5,6,7

4 ShareLock CREATE INDEX 3,5,6,7

5 ShareRowExclusiveLock 3,4,5,6,7

6 ExclusiveLock 2,3,4,5,6,7

7 AccessExclusiveLock DROP TABLE, ALTER TABLE, VACUUM all

All lock types can be obtained by user LOCK TABLE commands.

Locks are held till end of transaction: you can grab a lock, but you can’trelease it
except by ending your transaction.

18
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Lock implementation

Locks are recorded in a shared-memory hash table keyed by kind and ID ofobject
being locked. Each item shows the types and numbers of locks held orpending on
its object. Would-be lockers who have a conflict with an existinglock must wait.

Waiting is handled by waiting on a per-process IPC semaphore, which willbe

signaled when another process releases the wanted lock. Note we needonly one
semaphore per concurrent backend, not one per potentially lockableobject.

19
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Deadlock detection

Deadlock is possible if two transactions try to grab conflicting locksin different orders.

If a would-be locker sleeps for more than a second without getting thedesired lock,
it runs a deadlock-check algorithm that searches thelock hash table for circular
lock dependencies. If it finds any, thenobtaining the lock will be impossible, so it
gives up and reports anerror. Else it goes back to sleep and waits till granted the
lock (ortill client application gives up and requests transaction cancel).

• The delay before running the deadlock check algorithm can betuned to match the
typical transaction time in a particular server’sworkload. In this way, unnecessary
deadlock checks are seldomperformed, but real deadlocks are detected reasonably
quickly.

20
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Short-term locks

Short-term locks protect datastructures in shared memory, such as the lock

hashtable described above.

These locks should only be held for long enough toexamine and/or update a
shared item --- in particular a backend should neverblock while holding one.

Implementation: spin locks based on platform-specificatomic test-and-set

instructions. This allows the lock code to fall through extremely quicklyin the
common case where there is no contention for the lock. If the test-and-setfails,
we sleep for a short period (using select(2)) andtry again. No deadlock detection
as such, but we give up and report errorif fail too many times.

21
30 Oct 2000 Tom Lane
Transaction Processing in PostgreSQL

Summary

PostgreSQL offers true ACID semantics for transactions, given somereasonable

assumptions about the behavior of the underlying Unix kerneland hardware.

Multi-version concurrency control allows concurrent readingand writing of tables,

blocking only for concurrent updates of same row.

MVCC is practical because of non-overwriting storage manager that weinherited

from the Berkeley POSTQUEL project. Traditionalrow-overwriting storage
management would have a much harder time.

22
30 Oct 2000 Tom Lane