Does Postgres rewrite entire row on update?

Question

We run Postgres 9.0 on Windows 2008 Server. There is a large table contains a bytea column for storing binary data ranging from 0-5MB in each row:

CREATE TABLE files
(
  file_id serial NOT NULL,
  data bytea NOT NULL,
  on_disk boolean,
  CONSTRAINT files_pkey PRIMARY KEY (file_id)
)

Recently we have been updating the on_disk field for each row (not touching the data field). We believe this has eaten up space in our temp tablespace (or something), for two reasons:

1) We started receiving this error in other random parts of the system running large queries:

ERROR: 53100: could not write block 92271 of temporary file

2) Our free space dropped from ~7GB to 1.5GB within a week which is uncharacteristic.

Can anyone confirm:

a) Will updating a row in postgres cause it rewrite the ENTIRE row (including large binary data) without freeing up the old space? This would explain our symptoms

b) Is it writing to some other temporary tablespace during the change which also uses up space? (Can we force freeing up the temp space?)

c) Is there a way we can perform minor boolean field updates to this table WITHOUT rewriting the row (&chewing up diskspace) each time?

d) Can we force postgres periodically to free up the used space without rewriting the entire table?? (Our known methods for freeing up the space involve a table rewrite which we don't have the space for)

P.S.: Yes we are migrating our server to a host with a larger storage space... this may be 1-2 months away.

Answer 1

At least on 9.3, PostgreSQL does not rewrite fields stored out of line in TOAST tables if they are stored out of line. I don't know if that's true in 9.0.

You can see what storage is used for a column with \d+ tablename; the storage column shows the mode used. Individual tuples may be stored compressed in-line if they're small enough (ex: < 2K), even in an extended storage column where tuples are eligible for out of line storage.

See the documentation for TOAST and ALTER TABLE ... SET STORAGE.

Temp files are stored in the temp_tablespaces. By default this is empty, in which case it falls back to default_tablespace, which in turn if empty falls back to the pg_default tablespace.

Space within tables/indexes should be freed for re-use automatically by autovacuum. Make sure your autovacuum daemon is running often enough and doesn't have too much of a cost_delay set. Autovacuum has been significantly improved since 9.0.

If you want to free space back to the operating system or for use in other tables, you'll need to VACUUM FULL or use an external tool like pg_repack to do it in a less intrusive manner.

Answer 2

Picking c) from your questions:

Is there a way we can perform minor boolean field updates to this table WITHOUT rewriting the row (&chewing up diskspace) each time?

As @Craig already explained, columns that are "TOAST-able" and bigger than a certain threshold are stored out-of-line in a dedicated TOAST table per table (separate "relation forks", separate files on disk). So, a 5 MB bytea column would stay mostly untouched in an update if the column itself is not changed. The manual:

During an UPDATE operation, values of unchanged fields are normally preserved as-is; so an UPDATE of a row with out-of-line values incurs no TOAST costs if none of the out-of-line values change.

Bold emphasis mine.
The row in the main relation fork is still copied and a dead row stays behind when updated (whether or not any values actually changed). For large row sizes, the following solution might pay:

Create a small separate 1:1 table for frequently changed flags. Just the primary key (= foreign key at the same time) and the frequently changed flags. This would make updates a lot faster and preserve disk space - for an initial extra overhead and some cost for queries that need to join both tables (other queries actually get faster). More about on-disk space requirement of table rows:

• Making sense of Postgres row sizes