PostgreSQL – Memory Requirements for CREATE INDEX

indexmemorypostgresqlpostgresql-9.3

I've got 6GB of RAM on my laptop and for a test I'm creating a 50mil-rows table in PostgreSQL 9.3. I then want to create an index on the table.

The table and the resulting index together (or twice the table total size) can fit into 5GB of RAM and I set maintenance_work_mem to 5GB, still CREATE INDEX uses external sort with about 1.4GB of temp files. Why is that so?

Is my expectation that it should be able to sort in RAM unreasonable?

test=# set maintenance_work_mem to '5GB';
SET
test=# create table t1 as (select i::int, random() as f from generate_series(1, 50000000) i);
SELECT 50000000
test=# select pg_size_pretty(pg_relation_size('t1'));
 pg_size_pretty 
----------------
 2111 MB
(1 row)

test=# create index on t1(f, i);
CREATE INDEX
test=# select pg_size_pretty(pg_relation_size('t1_f_i_idx'));
 pg_size_pretty 
----------------
 1504 MB
(1 row)

In the server log:

LOG:  temporary file: path "base/pgsql_tmp/pgsql_tmp22623.1", size 1073741824
STATEMENT:  create index on t1(f, i);
LOG:  temporary file: path "base/pgsql_tmp/pgsql_tmp22623.2", size 327622656
LOG:  external sort ended, 171065 disk blocks used: CPU 6.78s/268.73u sec elapsed 313.18 sec

Is there a way to calculate CREATE INDEX memory requirement before actually running it?

Best Answer

Through version 9.3, the indirection array used for sorting had to fit in a single 1GB memory allocation. This created an artificial limit on the number of tuples which could be sorted in memory. Once that limit was reached, it had to switch to a disk sort, even if there was memory left over.

This restriction was removed in version 9.4.

OBSERVATION #1

Performance implications should become quickly apparent if the index pages of monitor.result have to experience two things:

Experience 1) Swapping of the RAM making up the MyISAM Key Cache
Experience 2) Rotation in and out of the MyISAM Key Cache (sized by key_buffer_size)

Experience 1 is virtually unavoidable. As for Experience #2, it could result in needed index pages being purged out of the MyISAM Key Cache in the face of more recent queries against other MyISAM tables. Those needed index pages more be brought back by querying the corresponding table. The two experiences togther could make for a slower-than-expected query on relatively small tables.

However, you could minimize or neutralize any ill effects of swapping by assigning the index of monitor.result by creating a dedicated MyISAM Key Cache. It will be a cache that will contain only index pages from monitor.result.

How do you do that ???

First, recall that you mentioned that the index usage for monitor.result was 2.29MB. You can create that dedicated key cache with that size with a little headroom, say 2.5MB. Let us create that key cache like this:

SET GLOBAL monitor_result_private_cache.key_buffer_size = 1024 * 512 * 5;
CACHE INDEX monitor.result IN monitor_result_private_cache;
LOAD INDEX INTO CACHE monitor.result;

This will perform the following steps:

Create the Key Cache
Assign the Key Cache to the MyISAM Table using LOAD INDEX INTO CACHE
Load the Index Pages of the Assigned MyISAM Table into its Corresponding Cache

It would conveniently keep the index pages of that table from ever leaving the cache. The only table index pages would leave is if INSERTs into monitor.result increases the contents beyond 2.5MB. You must choose enough headroom to accommodate many INSERTs into monitor.result.

OBSERVATION #2

I also noticed the index you laid out for monitor.result:

PRIMARY KEY (`server`, `request`, `recorded`, `ref_id`)

If any of your queries against monitor.result resemble something like this:

SELECT resoultion,value FROM monitor.result
WHERE server = 200 AND refid = 50 AND ... ;

You can speed queries by reordering the PRIMARY KEY

CREATE  TABLE IF NOT EXISTS `monitor`.`result` (            
  `server` SMALLINT UNSIGNED NOT NULL ,            
  `ref_id` SMALLINT UNSIGNED NOT NULL ,            
  `request` TINYINT UNSIGNED NOT NULL ,            
  `recorded` TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP ,            
  `resolution` TINYINT NOT NULL ,            
  `value` MEDIUMINT UNSIGNED NOT NULL ,            
  PRIMARY KEY (`server`, `ref_id`, `request`, `recorded`) )            
ENGINE = MyISAM

or adding a UNIQUE index

CREATE  TABLE IF NOT EXISTS `monitor`.`result` (             
  `server` SMALLINT UNSIGNED NOT NULL ,             
  `ref_id` SMALLINT UNSIGNED NOT NULL ,             
  `request` TINYINT UNSIGNED NOT NULL ,             
  `recorded` TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP ,             
  `resolution` TINYINT NOT NULL ,             
  `value` MEDIUMINT UNSIGNED NOT NULL ,             
  PRIMARY KEY (`server`, `request`, `recorded`, `ref_id`),
  UNIQUE KEY uniqndx1 (`server`, `ref_id`, `request`, `recorded`)
ENGINE = MyISAM

If you add a UNIQUE index, you must double the dedicate keycache accordingly.

OBSERVATION #3

You mentioned a SATA drive. Good choice for archival, low-update historical data. Any MyISAM table on a SATA drive which has a dedicated keycache should not be bothered by the index lookup but the data retrival time from the drive would be up to you to benchmark to see if you could live with the running times.

Here is an alternative:

Create an index that has every column

CREATE  TABLE IF NOT EXISTS `monitor`.`result` (            
  `server` SMALLINT UNSIGNED NOT NULL ,            
  `ref_id` SMALLINT UNSIGNED NOT NULL ,            
  `request` TINYINT UNSIGNED NOT NULL ,            
  `recorded` TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP ,            
  `resolution` TINYINT NOT NULL ,            
  `value` MEDIUMINT UNSIGNED NOT NULL ,            
  PRIMARY KEY (`server`, `ref_id`, `request`, `recorded`, `resolution`, `value`) )            
ENGINE = MyISAM

What does this do? It provides data retrieval of entire rows strictly from the index. Combining this with a dedicated keycache, you will essentially has the whole table in RAM. All queries would be fulfilled by the index and never touch the table, REGARDLESS of the table being on SAS, SATA, SSD, or even stone.

UPDATE 2012-01-26 18:18 EDT

Question 1: You may want to look into memcached. I believe that there is a version InnoDB with a memcached plugin. At least, that's what I heard.

Question 2: InnoDB is for transactional tables. If you have archive data, compressed MyISAM tables should fill the bill. In fact, you could look into the ARCHIVE storage engine.

Question 3: Storing an index on disk (MyISAM,InnoDB) is always standard and cannot be changed. You have use special commands or run special queries to preload caches.

Question 4: RAID-10 for high-writes, SSD for high-reads. Watch your disk surface temperatures !!!

Question 5: if the table is strictly for holding historical info, no need for overkill. As long as it is a table rarely read, no need for special caching considerations.

Best Answer

Related Solutions

Postgresql – create index – invalid memory alloc request size

Less RAM Than Index_Length in MyISAM – MySQL Optimization

OBSERVATION #1

OBSERVATION #2

OBSERVATION #3

UPDATE 2012-01-26 18:18 EDT

Related Question