release.sgml

PostgreSQL 8.1.4的源码 适用于Linux下的开源数据库系统

      <para>
       Currently <productname>PostgreSQL</productname> processes a
       backslash in a string literal as introducing a special escape sequence,
       e.g. <literal>\n</> or <literal>\010</>.
       While this allows easy entry of special values, it is
       nonstandard and makes porting of applications from other
       databases more difficult. For this reason, the
       <productname>PostgreSQL</productname> project is planning to
       remove the special meaning of backslashes in strings. For
       backward compatibility and for users who want special backslash
       processing, a new string syntax has been created. This new string
       syntax is formed by writing an <literal>E</> immediately preceding the
       single quote that starts the string, e.g. <literal>E'hi\n'</>. While
       this release does not change the handling of backslashes in strings, it
       does add new configuration parameters to help users migrate applications
       for future releases:
      </para>
      <itemizedlist>

       <listitem>
        <para>
         <varname>standard_conforming_strings</> &mdash; does this release
         treat backslashes literally in ordinary strings?
        </para>
       </listitem>
  
       <listitem>
       <para>
        <varname>escape_string_warning</> &mdash; warn about backslashes in
        ordinary (non-E) strings
       </para>
      </listitem>

      </itemizedlist>

      <para>
       The <varname>standard_conforming_strings</> value is read-only.
       Applications can retrieve the value to know how backslashes are
       processed.  (Presence of the parameter can also be taken as an
       indication that <literal>E''</> string syntax is supported.)
       In a future release, <varname>standard_conforming_strings</>
       will be true, meaning backslashes will be treated literally in
       non-E strings. To prepare for this change, use <literal>E''</>
       strings in places that need special backslash processing, and
       turn on <varname>escape_string_warning</> to find additional
       strings that need to be converted to use <literal>E''</>.
       Also, use two single-quotes (<literal>''</>) to embed a literal
       single-quote in a string, rather than the
       <productname>PostgreSQL</productname>-supported syntax of
       backslash single-quote (<literal>\'</>).  The former is
       standards-conforming and does not require the use of the
       <literal>E''</> string syntax.  You can also use the
       <literal>$$</> string syntax, which does not treat backslashes
       specially.
      </para>
     </listitem>

     <listitem>
      <para>
       Make <command>REINDEX DATABASE</> reindex all indexes in the
       database (Tom)
      </para>
      <para>
       Formerly, <command>REINDEX DATABASE</> reindexed only
       system tables. This new behavior seems more intuitive. A new
       command <command>REINDEX SYSTEM</> provides the old functionality
       of reindexing just the system tables.
      </para>
     </listitem>

     <listitem>
      <para>
       Read-only large object descriptors now obey MVCC snapshot semantics
      </para>
      <para>
       When a large object is opened with <literal>INV_READ</> (and not
       <literal>INV_WRITE</>), the data read from the descriptor will now
       reflect a <quote>snapshot</> of the large object's state at the
       time of the transaction snapshot in use by the query that called
       <function>lo_open()</>.  To obtain the old behavior of always
       returning the latest committed data, include <literal>INV_WRITE</>
       in the mode flags for <function>lo_open()</>.
      </para>
     </listitem>

     <listitem>
      <para>
       Add proper dependencies for arguments of sequence functions (Tom)
      </para>
      <para>
       In previous releases, sequence names passed to <function>nextval()</>,
       <function>currval()</>, and <function>setval()</> were stored as
       simple text strings, meaning that renaming or dropping a
       sequence used in a <literal>DEFAULT</> clause made the clause
       invalid. This release stores all newly-created sequence function
       arguments as internal OIDs, allowing them to track sequence
       renaming, and adding dependency information that prevents
       improper sequence removal. It also makes such <literal>DEFAULT</>
       clauses immune to schema renaming and search path changes.
      </para>
      <para>
       Some applications might rely on the old behavior of
       run-time lookup for sequence names. This can still be done by
       explicitly casting the argument to <type>text</>, for example
       <literal>nextval('myseq'::text)</>.
      </para>
      <para>
       Pre-8.1 database dumps loaded into 8.1 will use the old text-based
       representation and therefore will not have the features of
       OID-stored arguments. However, it is possible to update a
       database containing text-based <literal>DEFAULT</> clauses.
       First, save this query into a file, such as <filename>fixseq.sql</>:
<programlisting>
SELECT  'ALTER TABLE ' ||
    pg_catalog.quote_ident(n.nspname) || '.' ||
    pg_catalog.quote_ident(c.relname) ||
    ' ALTER COLUMN ' || pg_catalog.quote_ident(a.attname) ||
    ' SET DEFAULT ' ||
    regexp_replace(d.adsrc,
                   $$val\(\(('[^']*')::text\)::regclass$$,
                   $$val(\1$$,
                   'g') ||
    ';'
FROM    pg_namespace n, pg_class c, pg_attribute a, pg_attrdef d
WHERE   n.oid = c.relnamespace AND
    c.oid = a.attrelid AND
    a.attrelid = d.adrelid AND
    a.attnum = d.adnum AND
    d.adsrc ~ $$val\(\('[^']*'::text\)::regclass$$;
</programlisting>
       Next, run the query against a database to find what
       adjustments are required, like this for database <literal>db1</>:
<programlisting>
psql -t -f fixseq.sql db1
</programlisting>
       This will show the <command>ALTER TABLE</> commands needed to
       convert the database to the newer OID-based representation.
       If the commands look reasonable, run this to update the database:
<programlisting>
psql -t -f fixseq.sql db1 | psql -e db1
</programlisting>
       This process must be repeated in each database to be updated.
      </para>
     </listitem>

     <listitem>
      <para>
       In <application>psql</application>, treat unquoted
       <literal>\{digit}+</> sequences as octal (Bruce)
      </para>
      <para>
       In previous releases, <literal>\{digit}+</> sequences were
       treated as decimal, and only <literal>\0{digit}+</> were treated
       as octal. This change was made for consistency.
      </para>
     </listitem>

     <listitem>
      <para>
       Remove grammar productions for prefix and postfix <literal>%</>
       and <literal>^</> operators
       (Tom)
      </para>
      <para>
       These have never been documented and complicated the use of the
       modulus operator (<literal>%</>) with negative numbers.
      </para>
     </listitem>

     <listitem>
      <para>
       Make <literal>&amp;&lt;</> and <literal>&amp;&gt;</> for polygons
       consistent with the box "over" operators (Tom)
      </para>
     </listitem>

     <listitem>
      <para>
       <command>CREATE LANGUAGE</> may ignore the provided arguments
       in favor of information from <structname>pg_pltemplate</>
       (Tom)
      </para>
      <para>
       A new system catalog <structname>pg_pltemplate</> has been defined
       to carry information about the preferred definitions of procedural
       languages (such as whether they have validator functions).  When
       an entry exists in this catalog for the language being created,
       <command>CREATE LANGUAGE</> will ignore all its parameters except the
       language name and instead use the catalog information.  This measure
       was taken because of increasing problems with obsolete language
       definitions being loaded by old dump files.  As of 8.1,
       <application>pg_dump</> will dump procedural language definitions as
       just <command>CREATE LANGUAGE <replaceable>name</></command>, relying
       on a template entry to exist at load time.  We expect this will be a
       more future-proof representation.
      </para>
     </listitem>

     <listitem>
      <para>
       Make <function>pg_cancel_backend(int)</function> return a
       <type>boolean</type> rather than an <type>integer</type> (Neil)
      </para>
     </listitem>

     <listitem>
      <para>
       Some users are having problems loading UTF-8 data into 8.1.X.
       This is because previous versions allowed invalid UTF-8 byte
       sequences to be entered into the database, and this release
       properly accepts only valid UTF-8 sequences. One way to correct a
       dumpfile is to run the command <command>iconv -c -f UTF-8 -t
       UTF-8 -o cleanfile.sql dumpfile.sql</>. The <literal>-c</> option
       removes invalid character sequences. A diff of the two files will
       show the sequences that are invalid. <command>iconv</> reads the
       entire input file into memory so it might be necessary to use
       <application>split</> to break up the dump into multiple smaller
       files for processing.
      </para>
     </listitem>

    </itemizedlist>
   </sect2>
 
   <sect2>
    <title>Additional Changes</title>
 
    <para>
     Below you will find a detailed account of the additional changes
     between <productname>PostgreSQL</productname> 8.1 and the
     previous major release.
    </para>
 
    <sect3>
     <title>Performance Improvements</title>
     <itemizedlist>

      <listitem>
       <para>
        Improve GiST and R-tree index performance (Neil)
       </para>
      </listitem>

      <listitem>
       <para>
        Improve the optimizer, including auto-resizing of hash joins
        (Tom)
       </para>
      </listitem>

      <listitem>
       <para>
        Overhaul internal API in several areas
       </para>
      </listitem>

      <listitem>
       <para>
        Change WAL record CRCs from 64-bit to 32-bit (Tom)
       </para>
       <para>
        We determined that the extra cost of computing 64-bit CRCs was
        significant, and the gain in reliability too marginal to justify it.
       </para>
      </listitem>

      <listitem>
       <para>
        Prevent writing large empty gaps in WAL pages (Tom)
       </para>
      </listitem>

      <listitem>
       <para>
        Improve spinlock behavior on SMP machines, particularly Opterons (Tom)
       </para>
      </listitem>

      <listitem>
       <para>
        Allow nonconsecutive index columns to be used in a multicolumn
        index (Tom)
       </para>
       <para>
        For example, this allows an index on columns a,b,c to be used in
        a query with <command>WHERE a = 4 and c = 10</>.
       </para>
      </listitem>

      <listitem>
       <para>
        Skip WAL logging for <command>CREATE TABLE AS</> /
        <command>SELECT INTO</> (Simon)
       </para>
       <para>
        Since a crash during <command>CREATE TABLE AS</> would cause the
        table to be dropped during recovery, there is no reason to WAL
        log as the table is loaded.  (Logging still happens if WAL
        archiving is enabled, however.)
       </para>
      </listitem>

      <listitem>
       <para>
        Allow concurrent GIST index access (Teodor, Oleg)
       </para>
      </listitem>

      <listitem>
       <para>
        Add configuration parameter <varname>full_page_writes</> to
        control writing full pages to WAL (Bruce)
       </para>
       <para>
        To prevent partial disk writes from corrupting the database,
        <productname>PostgreSQL</productname> writes a complete copy of
        each database disk page to WAL the first time it is modified
        after a checkpoint. This option turns off that functionality for more
        speed.  This is safe to use with battery-backed disk caches where
        partial page writes cannot happen.
       </para>
      </listitem>

      <listitem>
       <para>
        Use <literal>O_DIRECT</> if available when using
        <literal>O_SYNC</> for <varname>wal_sync_method</varname>
        (Itagaki Takahiro)
       </para>
       <para>
        <literal>O_DIRECT</> causes disk writes to bypass the kernel
        cache, and for WAL writes, this improves performance.
       </para>
      </listitem>

      <listitem>
       <para>
        Improve <command>COPY FROM</> performance (Alon Goldshuv)
       </para>
       <para>
        This was accomplished by reading <command>COPY</> input in
        larger chunks, rather than character by character.
       </para>
      </listitem>

      <listitem>
       <para>
        Improve the performance of <function>COUNT()</function>,
        <function>SUM</function>, <function>AVG()</function>,
        <function>STDDEV()</function>, and
        <function>VARIANCE()</function> (Neil, Tom)
       </para>
      </listitem>
     </itemizedlist>
    </sect3>

    <sect3>
     <title>Server Changes</title>
     <itemizedlist>

      <listitem>
       <para>
        Prevent problems due to transaction ID (XID) wraparound (Tom)
       </para>
       <para>
        The server will now warn when the transaction counter approaches
        the wraparound point.  If the counter becomes too close to wraparound,
        the server will stop accepting queries.  This ensures that data is
        not lost before needed vacuuming is performed.
       </para>
      </listitem>

      <listitem>
       <para>
        Fix problems with object IDs (OIDs) conflicting with existing system
        objects after the OID counter has wrapped around (Tom)
       </para>
      </listitem>
  
      <listitem>
       <para>
        Add warning about the need to increase
        <varname>max_fsm_relations</> and <varname>max_fsm_pages</>
        during <command>VACUUM</> (Ron Mayer)
       </para>
      </listitem>

      <listitem>
       <para>
        Add <varname>temp_buffers</> configuration parameter to allow
        users to determine the size of the local buffer area for
        temporary table access (Tom)
       </para>