On Mon, Nov 21, 2022 at 11:15:36AM +0100, Laurenz Albe wrote:
> > --- a/doc/src/sgml/ref/release_savepoint.sgml
> > +++ b/doc/src/sgml/ref/release_savepoint.sgml
>
> > + <command>RELEASE SAVEPOINT</command> releases the named savepoint and
> > + all active savepoints that were created after the named savepoint,
> > + and frees their resources. All changes made since the creation of the
> > + savepoint, excluding rolled back savepoints changes, are merged into
> > + the transaction or savepoint that was active when the named savepoint
> > + was created. Changes made after <command>RELEASE SAVEPOINT</command>
> > + will also be part of this active transaction or savepoint.
>
> I am not sure if "rolled back savepoints changes" is clear enough.
> I understand that you are trying to avoid the term "subtransaction".
> How about:
>
> All changes made since the creation of the savepoint that didn't already
> get rolled back are merged ...
Yes, I like that, changed.
> > --- a/doc/src/sgml/ref/rollback.sgml
> > +++ b/doc/src/sgml/ref/rollback.sgml
> >
> > + If <literal>AND CHAIN</literal> is specified, a new (unaborted)
>
> *Sigh* I'll make one last plea for "not aborted".
Uh, I thought you wanted "unaborted", but I now changed it to "not
aborted".
> > --- /dev/null
> > +++ b/doc/src/sgml/xact.sgml
>
> > + <para>
> > + Transactions can be created explicitly using <command>BEGIN</command>
> > + and <command>COMMIT</command>, which creates a transaction block.
> > + An SQL statement outside of a transaction block automatically uses
> > + a single-statement transaction.
> > + </para>
>
> Sorry, I should have noted that the first time around.
>
> Transactions are not created using COMMIT.
>
> Perhaps we should also avoid the term "transaction block". Even without
> speaking
> of a "block", way too many people confuse PL/pgSQL's BEGIN ... END blocks
> with transactions. On the other hand, we use "transaction block" everywhere
> else in the documentation...
>
> How about:
>
> <para>
> Multi-statement transactions can be created explicitly using
> <command>BEGIN</command> or <command>START TRANSACTION</command> and
> are ended using <command>COMMIT</command> or <command>ROLLBACK</command>.
> An SQL statement outside of a transaction block automatically uses
> a single-statement transaction.
> </para>
I used your wording, but technically you can use BEGIN/COMMIT with a
single statement, so multi-statement it not a requirement, so I used
your text but removed "Multi-statement":
Transactions can be created explicitly using <command>BEGIN</command> or
<command>START TRANSACTION</command> and ended using
<command>COMMIT</command> or <command>ROLLBACK</command>.
> > + <sect1 id="xact-locking">
> > +
> > + <title>Transactions and Locking</title>
> > +
> > + <para>
> > + The transaction IDs of currently executing transactions are shown in
> > + <link linkend="view-pg-locks"><structname>pg_locks</structname></link>
> > + in columns <structfield>virtualxid</structfield> and
> > + <structfield>transactionid</structfield>. Read-only transactions
> > + will have <structfield>virtualxid</structfield>s but NULL
> > + <structfield>transactionid</structfield>s, while read-write transactions
> > + will have both as non-NULL.
> > + </para>
>
> Perhaps the following will be prettier than "have both as non-NULL":
>
> ..., while both columns will be set in read-write transactions.
Agreed, changed. Updated patch attached.
--
Bruce Momjian <br...@momjian.us> https://momjian.us
EDB https://enterprisedb.com
Indecision is a decision. Inaction is an action. Mark Batterson
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 24b1624bad..01c65ede4c 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -7229,12 +7229,14 @@ local0.* /var/log/postgresql
</row>
<row>
<entry><literal>%v</literal></entry>
- <entry>Virtual transaction ID (backendID/localXID)</entry>
+ <entry>Virtual transaction ID (backendID/localXID); see
+ <xref linkend="transaction-id"/></entry>
<entry>no</entry>
</row>
<row>
<entry><literal>%x</literal></entry>
- <entry>Transaction ID (0 if none is assigned)</entry>
+ <entry>Transaction ID (0 if none is assigned); see
+ <xref linkend="transaction-id"/></entry>
<entry>no</entry>
</row>
<row>
diff --git a/doc/src/sgml/datatype.sgml b/doc/src/sgml/datatype.sgml
index b030b36002..fdffba4442 100644
--- a/doc/src/sgml/datatype.sgml
+++ b/doc/src/sgml/datatype.sgml
@@ -4992,7 +4992,8 @@ WHERE ...
<structfield>xmin</structfield> and <structfield>xmax</structfield>. Transaction identifiers are 32-bit quantities.
In some contexts, a 64-bit variant <type>xid8</type> is used. Unlike
<type>xid</type> values, <type>xid8</type> values increase strictly
- monotonically and cannot be reused in the lifetime of a database cluster.
+ monotonically and cannot be reused in the lifetime of a database
+ cluster. See <xref linkend="transaction-id"/> for more details.
</para>
<para>
diff --git a/doc/src/sgml/filelist.sgml b/doc/src/sgml/filelist.sgml
index de450cd661..0d6be9a2fa 100644
--- a/doc/src/sgml/filelist.sgml
+++ b/doc/src/sgml/filelist.sgml
@@ -104,6 +104,7 @@
<!ENTITY protocol SYSTEM "protocol.sgml">
<!ENTITY sources SYSTEM "sources.sgml">
<!ENTITY storage SYSTEM "storage.sgml">
+<!ENTITY transaction SYSTEM "xact.sgml">
<!ENTITY tablesample-method SYSTEM "tablesample-method.sgml">
<!ENTITY generic-wal SYSTEM "generic-wal.sgml">
<!ENTITY custom-rmgr SYSTEM "custom-rmgr.sgml">
diff --git a/doc/src/sgml/func.sgml b/doc/src/sgml/func.sgml
index 82fba48d5f..2a083eb015 100644
--- a/doc/src/sgml/func.sgml
+++ b/doc/src/sgml/func.sgml
@@ -24676,7 +24676,10 @@ SELECT collation for ('foo' COLLATE "de_DE");
<para>
Returns the current transaction's ID. It will assign a new one if the
current transaction does not have one already (because it has not
- performed any database updates).
+ performed any database updates); see <xref
+ linkend="transaction-id"/> for details. If executed in a
+ subtransaction, this will return the top-level transaction ID;
+ see <xref linkend="subxacts"/> for details.
</para></entry>
</row>
@@ -24693,6 +24696,8 @@ SELECT collation for ('foo' COLLATE "de_DE");
ID is assigned yet. (It's best to use this variant if the transaction
might otherwise be read-only, to avoid unnecessary consumption of an
XID.)
+ If executed in a subtransaction, this will return the top-level
+ transaction ID.
</para></entry>
</row>
@@ -24736,6 +24741,9 @@ SELECT collation for ('foo' COLLATE "de_DE");
<para>
Returns a current <firstterm>snapshot</firstterm>, a data structure
showing which transaction IDs are now in-progress.
+ Only top-level transaction ID are included in the snapshot;
+ subtransaction ID are not shown; see <xref linkend="subxacts"/>
+ for details.
</para></entry>
</row>
@@ -24790,7 +24798,8 @@ SELECT collation for ('foo' COLLATE "de_DE");
Is the given transaction ID <firstterm>visible</firstterm> according
to this snapshot (that is, was it completed before the snapshot was
taken)? Note that this function will not give the correct answer for
- a subtransaction ID.
+ a subtransaction ID (subxid); see <xref linkend="subxacts"/> for
+ details.
</para></entry>
</row>
</tbody>
@@ -24802,8 +24811,9 @@ SELECT collation for ('foo' COLLATE "de_DE");
wraps around every 4 billion transactions. However,
the functions shown in <xref linkend="functions-pg-snapshot"/> use a
64-bit type <type>xid8</type> that does not wrap around during the life
- of an installation, and can be converted to <type>xid</type> by casting if
- required. The data type <type>pg_snapshot</type> stores information about
+ of an installation and can be converted to <type>xid</type> by casting if
+ required; see <xref linkend="transaction-id"/> for details.
+ The data type <type>pg_snapshot</type> stores information about
transaction ID visibility at a particular moment in time. Its components
are described in <xref linkend="functions-pg-snapshot-parts"/>.
<type>pg_snapshot</type>'s textual representation is
@@ -24849,7 +24859,7 @@ SELECT collation for ('foo' COLLATE "de_DE");
xmax</literal> and not in this list was already completed at the time
of the snapshot, and thus is either visible or dead according to its
commit status. This list does not include the transaction IDs of
- subtransactions.
+ subtransactions (subxids).
</entry>
</row>
</tbody>
diff --git a/doc/src/sgml/glossary.sgml b/doc/src/sgml/glossary.sgml
index 93fb149d9a..7c01a541fe 100644
--- a/doc/src/sgml/glossary.sgml
+++ b/doc/src/sgml/glossary.sgml
@@ -1777,7 +1777,8 @@
<literal>3</literal> (values under that are reserved) and the
epoch value is incremented by one.
In some contexts, the epoch and xid values are
- considered together as a single 64-bit value.
+ considered together as a single 64-bit value; see <xref
+ linkend="transaction-id"/> for more details.
</para>
<para>
For more information, see
diff --git a/doc/src/sgml/monitoring.sgml b/doc/src/sgml/monitoring.sgml
index e2426f7210..186aaec304 100644
--- a/doc/src/sgml/monitoring.sgml
+++ b/doc/src/sgml/monitoring.sgml
@@ -918,7 +918,8 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser
<structfield>backend_xid</structfield> <type>xid</type>
</para>
<para>
- Top-level transaction identifier of this backend, if any.
+ Top-level transaction identifier of this backend, if any; see
+ <xref linkend="transaction-id"/>.
</para></entry>
</row>
@@ -1890,7 +1891,8 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser
</row>
<row>
<entry><literal>virtualxid</literal></entry>
- <entry>Waiting to acquire a virtual transaction ID lock.</entry>
+ <entry>Waiting to acquire a virtual transaction ID lock; see
+ <xref linkend="transaction-id"/>.</entry>
</row>
</tbody>
</tgroup>
diff --git a/doc/src/sgml/pgrowlocks.sgml b/doc/src/sgml/pgrowlocks.sgml
index 2914bf6e6d..ad15cda668 100644
--- a/doc/src/sgml/pgrowlocks.sgml
+++ b/doc/src/sgml/pgrowlocks.sgml
@@ -57,7 +57,8 @@ pgrowlocks(text) returns setof record
<row>
<entry><structfield>locker</structfield></entry>
<entry><type>xid</type></entry>
- <entry>Transaction ID of locker, or multixact ID if multitransaction</entry>
+ <entry>Transaction ID of locker, or multixact ID if
+ multitransaction; see <xref linkend="transaction-id"/></entry>
</row>
<row>
<entry><structfield>multi</structfield></entry>
diff --git a/doc/src/sgml/postgres.sgml b/doc/src/sgml/postgres.sgml
index 73439c049e..2e271862fc 100644
--- a/doc/src/sgml/postgres.sgml
+++ b/doc/src/sgml/postgres.sgml
@@ -271,6 +271,7 @@ break is not needed in a wider output rendering.
&brin;
&hash;
&storage;
+ &transaction;
&bki;
&planstats;
&backup-manifest;
diff --git a/doc/src/sgml/ref/commit.sgml b/doc/src/sgml/ref/commit.sgml
index 5f244cdd3c..53d830998c 100644
--- a/doc/src/sgml/ref/commit.sgml
+++ b/doc/src/sgml/ref/commit.sgml
@@ -62,6 +62,9 @@ COMMIT [ WORK | TRANSACTION ] [ AND [ NO ] CHAIN ]
linkend="sql-set-transaction"/>) as the just finished one. Otherwise,
no new transaction is started.
</para>
+ <para>
+ The SQL Standard describes this as a chained transaction.
+ </para>
</listitem>
</varlistentry>
</variablelist>
diff --git a/doc/src/sgml/ref/release_savepoint.sgml b/doc/src/sgml/ref/release_savepoint.sgml
index daf8eb9a43..4adb422204 100644
--- a/doc/src/sgml/ref/release_savepoint.sgml
+++ b/doc/src/sgml/ref/release_savepoint.sgml
@@ -21,7 +21,7 @@ PostgreSQL documentation
<refnamediv>
<refname>RELEASE SAVEPOINT</refname>
- <refpurpose>destroy a previously defined savepoint</refpurpose>
+ <refpurpose>release a previously defined savepoint</refpurpose>
</refnamediv>
<refsynopsisdiv>
@@ -34,23 +34,13 @@ RELEASE [ SAVEPOINT ] <replaceable>savepoint_name</replaceable>
<title>Description</title>
<para>
- <command>RELEASE SAVEPOINT</command> destroys a savepoint previously defined
- in the current transaction.
- </para>
-
- <para>
- Destroying a savepoint makes it unavailable as a rollback point,
- but it has no other user visible behavior. It does not undo the
- effects of commands executed after the savepoint was established.
- (To do that, see <xref linkend="sql-rollback-to"/>.)
- Destroying a savepoint when
- it is no longer needed allows the system to reclaim some resources
- earlier than transaction end.
- </para>
-
- <para>
- <command>RELEASE SAVEPOINT</command> also destroys all savepoints that were
- established after the named savepoint was established.
+ <command>RELEASE SAVEPOINT</command> releases the named savepoint and
+ all active savepoints that were created after the named savepoint,
+ and frees their resources. All changes made since the creation of
+ the savepoint that didn't already get rolled back are merged into
+ the transaction or savepoint that was active when the named savepoint
+ was created. Changes made after <command>RELEASE SAVEPOINT</command>
+ will also be part of this active transaction or savepoint.
</para>
</refsect1>
@@ -62,7 +52,7 @@ RELEASE [ SAVEPOINT ] <replaceable>savepoint_name</replaceable>
<term><replaceable>savepoint_name</replaceable></term>
<listitem>
<para>
- The name of the savepoint to destroy.
+ The name of the savepoint to release.
</para>
</listitem>
</varlistentry>
@@ -78,7 +68,7 @@ RELEASE [ SAVEPOINT ] <replaceable>savepoint_name</replaceable>
<para>
It is not possible to release a savepoint when the transaction is in
- an aborted state.
+ an aborted state; to do that, use <xref linkend="sql-rollback-to"/>.
</para>
<para>
@@ -93,7 +83,7 @@ RELEASE [ SAVEPOINT ] <replaceable>savepoint_name</replaceable>
<title>Examples</title>
<para>
- To establish and later destroy a savepoint:
+ To establish and later release a savepoint:
<programlisting>
BEGIN;
INSERT INTO table1 VALUES (3);
@@ -104,6 +94,36 @@ COMMIT;
</programlisting>
The above transaction will insert both 3 and 4.
</para>
+
+ <para>
+ A more complex example with multiple nested subtransactions:
+<programlisting>
+BEGIN;
+ INSERT INTO table1 VALUES (1);
+ SAVEPOINT sp1;
+ INSERT INTO table1 VALUES (2);
+ SAVEPOINT sp2;
+ INSERT INTO table1 VALUES (3);
+ RELEASE SAVEPOINT sp2;
+ INSERT INTO table1 VALUES (4))); -- generates an error
+</programlisting>
+ In this example, the application requests the release of the savepoint
+ <literal>sp2</literal>, which inserted 3. This changes the insert's
+ transaction context to <literal>sp1</literal>. When the statement
+ attempting to insert value 4 generates an error, the insertion of 2 and
+ 4 are lost because they are in the same, now-rolled back savepoint,
+ and value 3 is in the same transaction context. The application can
+ now only choose one of these two commands, since all other commands
+ will be ignored with a warning:
+<programlisting>
+ ROLLBACK;
+ ROLLBACK TO SAVEPOINT sp1;
+</programlisting>
+ Choosing <command>ROLLBACK</command> will abort everything, including
+ value 1, whereas <command>ROLLBACK TO SAVEPOINT sp1</command> will retain
+ value 1 and allow the transaction to continue.
+ </para>
+
</refsect1>
<refsect1>
diff --git a/doc/src/sgml/ref/rollback.sgml b/doc/src/sgml/ref/rollback.sgml
index 142f71e774..7700547669 100644
--- a/doc/src/sgml/ref/rollback.sgml
+++ b/doc/src/sgml/ref/rollback.sgml
@@ -56,10 +56,10 @@ ROLLBACK [ WORK | TRANSACTION ] [ AND [ NO ] CHAIN ]
<term><literal>AND CHAIN</literal></term>
<listitem>
<para>
- If <literal>AND CHAIN</literal> is specified, a new transaction is
- immediately started with the same transaction characteristics (see <xref
- linkend="sql-set-transaction"/>) as the just finished one. Otherwise,
- no new transaction is started.
+ If <literal>AND CHAIN</literal> is specified, a new (not aborted)
+ transaction is immediately started with the same transaction
+ characteristics (see <xref linkend="sql-set-transaction"/>) as the
+ just finished one. Otherwise, no new transaction is started.
</para>
</listitem>
</varlistentry>
diff --git a/doc/src/sgml/ref/rollback_to.sgml b/doc/src/sgml/ref/rollback_to.sgml
index 27fa95cd1b..32c1bb9723 100644
--- a/doc/src/sgml/ref/rollback_to.sgml
+++ b/doc/src/sgml/ref/rollback_to.sgml
@@ -35,8 +35,9 @@ ROLLBACK [ WORK | TRANSACTION ] TO [ SAVEPOINT ] <replaceable>savepoint_name</re
<para>
Roll back all commands that were executed after the savepoint was
- established. The savepoint remains valid and can be rolled back to
- again later, if needed.
+ established and then start a new subtransaction at the same transaction level.
+ The savepoint remains valid and can be rolled back to again later,
+ if needed.
</para>
<para>
diff --git a/doc/src/sgml/system-views.sgml b/doc/src/sgml/system-views.sgml
index 7c716fe327..4d8bc659f2 100644
--- a/doc/src/sgml/system-views.sgml
+++ b/doc/src/sgml/system-views.sgml
@@ -1416,7 +1416,8 @@
</para>
<para>
Virtual ID of the transaction targeted by the lock,
- or null if the target is not a virtual transaction ID
+ or null if the target is not a virtual transaction ID; see
+ <xref linkend="transactions"/>
</para></entry>
</row>
@@ -1425,8 +1426,8 @@
<structfield>transactionid</structfield> <type>xid</type>
</para>
<para>
- ID of the transaction targeted by the lock,
- or null if the target is not a transaction ID
+ ID of the transaction targeted by the lock, or null if the target
+ is not a transaction ID; <xref linkend="transactions"/>
</para></entry>
</row>
diff --git a/doc/src/sgml/wal.sgml b/doc/src/sgml/wal.sgml
index 6a38b53744..ed7929cbcd 100644
--- a/doc/src/sgml/wal.sgml
+++ b/doc/src/sgml/wal.sgml
@@ -4,8 +4,9 @@
<title>Reliability and the Write-Ahead Log</title>
<para>
- This chapter explains how the Write-Ahead Log is used to obtain
- efficient, reliable operation.
+ This chapter explains how to control the reliability of
+ <productname>PostgreSQL</productname>, including details about the
+ Write-Ahead Log.
</para>
<sect1 id="wal-reliability">
@@ -909,4 +910,5 @@
seem to be a problem in practice.
</para>
</sect1>
+
</chapter>
diff --git a/doc/src/sgml/xact.sgml b/doc/src/sgml/xact.sgml
new file mode 100644
index 0000000000..e483db0d9a
--- /dev/null
+++ b/doc/src/sgml/xact.sgml
@@ -0,0 +1,206 @@
+<!-- doc/src/sgml/mvcc.sgml -->
+
+<chapter id="transactions">
+
+ <title>Transaction Processing</title>
+
+ <para>
+ This chapter provides an overview of the internals of
+ <productname>PostgreSQL</productname>'s transaction management system.
+ The word transaction is often abbreviated as <firstterm>xact</firstterm>.
+ </para>
+
+ <sect1 id="transaction-id">
+
+ <title>Transactions and Identifiers</title>
+
+ <para>
+ Transactions can be created explicitly using <command>BEGIN</command>
+ or <command>START TRANSACTION</command> and ended using
+ <command>COMMIT</command> or <command>ROLLBACK</command>. An SQL
+ statement outside of a transaction block automatically uses a
+ single-statement transaction.
+ </para>
+
+ <para>
+ Every transaction is identified by a unique
+ <literal>VirtualTransactionId</literal> (also called
+ <literal>virtualXID</literal> or <literal>vxid</literal>), which
+ is comprised of a backend ID (or <literal>backendID</literal>)
+ and a sequentially-assigned number local to each backend, known as
+ <literal>localXID</literal>. For example, the virtual transaction
+ ID <literal>4/12532</literal> has a <literal>backendID</literal>
+ of <literal>4</literal> and a <literal>localXID</literal> of
+ <literal>12532</literal>.
+ </para>
+
+ <para>
+ Non-virtual <literal>TransactionId</literal> (or <type>xid</type>),
+ e.g., <literal>278394</literal>, are assigned sequentially to
+ transactions from a global counter used by all databases within
+ the <productname>PostgreSQL</productname> cluster. This assignment
+ happens when a transaction first writes to the database. This means
+ lower-numbered xids started writing before higher-numbered xids.
+ Note that the order in which transactions perform their first database
+ write might be different from the order in which the transactions
+ started, particularly if the transaction started with statements that
+ only performed database reads.
+ </para>
+
+ <para>
+ The internal transaction ID type <type>xid</type> is 32 bits wide
+ and <link linkend="vacuum-for-wraparound">wraps around</link> every
+ 4 billion transactions. A 32-bit epoch is incremented during each
+ wraparound. There is also a 64-bit type <type>xid8</type> which
+ includes this epoch and therefore does not wrap around during the
+ life of an installation; it can be converted to xid by casting.
+ The functions in <xref linkend="functions-pg-snapshot"/>
+ return <type>xid8</type> values. Xids are used as the
+ basis for <productname>PostgreSQL</productname>'s <link
+ linkend="mvcc">MVCC</link> concurrency mechanism and streaming
+ replication.
+ </para>
+
+ <para>
+ When a top-level transaction with a (non-virtual) xid commits,
+ it is marked as committed in the <filename>pg_xact</filename>
+ directory. Additional information is recorded in the
+ <filename>pg_commit_ts</filename> directory if <xref
+ linkend="guc-track-commit-timestamp"/> is enabled.
+ </para>
+
+ <para>
+ In addition to <literal>vxid</literal> and <type>xid</type>,
+ prepared transactions are also assigned Global Transaction
+ Identifiers (<acronym>GID</acronym>). GIDs are string literals up
+ to 200 bytes long, which must be unique amongst other currently
+ prepared transactions. The mapping of GID to xid is shown in <link
+ linkend="view-pg-prepared-xacts"><structname>pg_prepared_xacts</structname></link>.
+ </para>
+ </sect1>
+
+ <sect1 id="xact-locking">
+
+ <title>Transactions and Locking</title>
+
+ <para>
+ The transaction IDs of currently executing transactions are shown in
+ <link linkend="view-pg-locks"><structname>pg_locks</structname></link>
+ in columns <structfield>virtualxid</structfield> and
+ <structfield>transactionid</structfield>. Read-only transactions
+ will have <structfield>virtualxid</structfield>s but NULL
+ <structfield>transactionid</structfield>s, while both columns will be
+ set in read-write transactions.
+ </para>
+
+ <para>
+ Some lock types wait on <structfield>virtualxid</structfield>,
+ while other types wait on <structfield>transactionid</structfield>.
+ Row-level read and write locks are recorded directly in the locked
+ rows and can be inspected using the <xref linkend="pgrowlocks"/>
+ extension. Row-level read locks might also require the assignment
+ of multixact IDs (<literal>mxid</literal>; see <xref
+ linkend="vacuum-for-multixact-wraparound"/>).
+ </para>
+ </sect1>
+
+ <sect1 id="subxacts">
+
+ <title>Subtransactions</title>
+
+ <para>
+ Subtransactions are started inside transactions, allowing large
+ transactions to be broken into smaller units. Subtransactions can
+ commit or abort without affecting their parent transactions, allowing
+ parent transactions to continue. This allows errors to be handled
+ more easily, which is a common application development pattern.
+ The word subtransaction is often abbreviated as
+ <firstterm>subxact</firstterm>.
+ </para>
+
+ <para>
+ Subtransactions can be started explicitly using the
+ <command>SAVEPOINT</command> command, but can also be started in
+ other ways, such as PL/pgSQL's <command>EXCEPTION</command> clause.
+ PL/Python and PL/TCL also support explicit subtransactions.
+ Subtransactions can also be started from other subtransactions.
+ The top-level transaction and its child subtransactions form a
+ hierarchy or tree, which is why we refer to the main transaction as
+ the top-level transaction.
+ </para>
+
+ <para>
+ If a subtransaction is assigned a non-virtual transaction ID,
+ its transaction ID is referred to as a <quote>subxid</quote>.
+ Read-only subtransactions are not assigned subxids, but once they
+ attempt to write, they will be assigned one. This also causes all of
+ a subxid's parents, up to and including the top-level transaction,
+ to be assigned non-virtual transaction ids. We ensure that a parent
+ xid is always lower than any of its child subxids.
+ </para>
+
+ <para>
+ The immediate parent xid of each subxid is recorded in the
+ <filename>pg_subtrans</filename> directory. No entry is made for
+ top-level xids since they do not have a parent, nor is an entry made
+ for read-only subtransactions.
+ </para>
+
+ <para>
+ When a subtransaction commits, all of its committed child
+ subtransactions with subxids will also be considered subcommitted
+ in that transaction. When a subtransaction aborts, all of its child
+ subtransactions will also be considered aborted.
+ </para>
+
+ <para>
+ When a top-level transaction with an xid commits, all of its
+ subcommitted child subtransactions are also persistently recorded
+ as committed in the <filename>pg_xact</filename> directory. If the
+ top-level transaction aborts, all its subtransactions are also aborted,
+ even if they were subcommitted.
+ </para>
+
+ <para>
+ The more subtransactions each transaction keeps open (not
+ rolled back or released), the greater the transaction management
+ overhead. Up to 64 open subxids are cached in shared memory for
+ each backend; after that point, the storage I/O overhead increases
+ significantly due to additional lookups of subxid entries in
+ <filename>pg_subtrans</filename>.
+ </para>
+ </sect1>
+
+ <sect1 id="two-phase">
+
+ <title>Two-Phase Transactions</title>
+
+ <para>
+ <productname>PostgreSQL</productname> supports a two-phase commit (2PC)
+ protocol that allows multiple distributed systems to work together
+ in a transactional manner. The commands are <command>PREPARE
+ TRANSACTION</command>, <command>COMMIT PREPARED</command> and
+ <command>ROLLBACK PREPARED</command>. Two-phase transactions
+ are intended for use by external transaction management systems.
+ <productname>PostgreSQL</productname> follows the features and model
+ proposed by the X/Open XA standard, but does not implement some less
+ often used aspects.
+ </para>
+
+ <para>
+ When the user executes <command>PREPARE TRANSACTION</command>, the
+ only possible next commands are <command>COMMIT PREPARED</command>
+ or <command>ROLLBACK PREPARED</command>. In general, this prepared
+ state is intended to be of very short duration, but external
+ availability issues might mean transactions stay in this state
+ for an extended interval. Short-lived prepared
+ transactions are stored only in shared memory and WAL.
+ Transactions that span checkpoints are recorded in the
+ <filename>pg_twophase</filename> directory. Transactions
+ that are currently prepared can be inspected using <link
+ linkend="view-pg-prepared-xacts"><structname>pg_prepared_xacts</structname></link>.
+ </para>
+ </sect1>
+
+</chapter>
+
