Adding index for meta_key and post_id in postmeta

[DuckDagobert]

We should add an index on post_id and meta_key as queries that contain these columns are incredibly common (even in loop in core on every page load basically).

An index will improve query speed up to 100% even when only 1000 posts are in db.

Test it yourself:

ALTER TABLE `wp_postmeta` ADD INDEX `wp_postmeta_key_id` (`meta_key`,`post_id`);

[mukesh27]

@DuckDagobert when we have to call above sql query in core software?

[DuckDagobert]

No, my code is the fix/patch for the problem.

For query example take the loop query (which gets post_id and _visibility for example).
But most WP core but also plugin queries like woocommerce,... will profit immensely from this index, as they select post_id inner join on postmeta where meta_key

[fliespl]

I can confirm that adding such might improve performance on many plugins.

Let's take imagify for example (with 100000 attachment in database).

Query:

SELECT SQL_NO_CACHE count(1)
FROM wp_posts AS p 
INNER JOIN wp_postmeta AS imrwpmt1 
ON ( p.ID = imrwpmt1.post_id
AND imrwpmt1.meta_key = '_wp_attached_file' ) 
INNER JOIN wp_postmeta AS imrwpmt2 
ON ( p.ID = imrwpmt2.post_id
AND imrwpmt2.meta_key = '_wp_attachment_metadata' ) 
INNER JOIN wp_postmeta AS mt1 
ON ( p.ID = mt1.post_id
AND mt1.meta_key = '_imagify_status' ) 
WHERE p.post_mime_type IN ( 'image/jpeg','image/png','image/gif','application/pdf' ) 
AND p.post_type = 'attachment' 
AND p.post_status IN ( 'inherit','private' ) 
AND mt1.meta_value IN ( 'success', 'already_optimized' ) 
AND imrwpmt1.meta_value NOT LIKE '%://%'
AND imrwpmt1.meta_value NOT LIKE '_:\\\%'
AND ( LOWER( imrwpmt1.meta_value ) LIKE '%.jpg'
OR LOWER( imrwpmt1.meta_value ) LIKE '%.jpeg'
OR LOWER( imrwpmt1.meta_value ) LIKE '%.jpe'
OR LOWER( imrwpmt1.meta_value ) LIKE '%.png'
OR LOWER( imrwpmt1.meta_value ) LIKE '%.gif'
OR LOWER( imrwpmt1.meta_value ) LIKE '%.pdf' );

a) default wordpress indexes:
1 row in set (28.22 sec)

b) after adding post_id + meta_key index

ALTER TABLE wp_postmeta ADD INDEX( `post_id`, `meta_key`);

1 row in set (2.51 sec)

It could be 10 times faster in some situations.

[lukefiretoss]

This index in the sites database would really help a number of WordPress sites.

[galbaras]

I'd say an index like this is beneficial if it helps frontend visitors, or if it helps backend users without causing any frontend delays. With the wide variety of plugins that use post meta, this may be difficult to determine, so maybe site admins should consider this on a site-by-site basis.

@fliespl @DuckDagobert Have you tested both "post_id, meta_key" and "meta_key, post_id" to see which one is faster? These columns vary in cardinality and speed of comparison, so one may be better than the other.

Also, have you measured the impact on both frontend and backend? Can you share more contextual details?

[programmin]

What are the potential downsides of this - why this has not been added in WP db upgrades?

[galbaras]

An additional index means slower updates and slightly slower queries that don't use the index (more decisions to make on the best query plan). On sites that update meta data a lot, or sites that don't use this index much, this may not be of benefit.

Also, as I've mentioned, the index can be created in 2 ways, and this will yield different results - one may help more and one may hinder more. It really depends on the context.

Looking at the very commonly-used core functions for getting meta data, they seem to fetch all the meta data of specific IDs into a cache at once, and then grab specific values from that cache by the key, so an index on the meta key will not help speed this up.

And index on post_id, meta_key would speed things up if WordPress fetched single records by the post ID and meta key.

Core Team, am I reading this correctly? Because this could be the start of a course change that will bring much performance improvement.

Importantly, if access to wp_postmeta is done by post_id, the table should have a primary index in that field, and it's currently indexed by meta_id, which doesn't help much.

A primary index by post_id should be lightning fast, because all the meta data records for a given ID would be contiguous and require the smallest possible number of block fetches in the database.

[johnbillion]

Additional indexes also take up disk space. Ironically the sites that would benefit the most from this additional index would also experience the greatest increase in space usage. (I'm not against the change, just pointing this out.)

[galbaras]

@johnbillion Any thoughts on making post_id the primary index (easier) or changing the way WordPress retrieves meta data (faster)?

Looks like disk space is now taken up by indexes that don't help (as much as they could), and some of those can be removed in favour of better ones.

[fliespl]

@galbaras I don't have any big performance results for this index, but post_id + meta_key (80) was a winner in my case.

I have used limitation of 80 length for meta_key to reduce index space usage + because 99% of meta_keys were well below that value, while testing on a medium sized site (100k+ posts).

​https://ss.codeone.pl/ss-2023-12-18-22-42-24-1702935744-a7HMU6OU.png

Most of the use cases for postmeta I have seen on many sites involved combination of post_id + meta_key (like filtering posts for specific ACF value or woocommerce value).

Also I disagree about post_id primary key (It's not even currently possible - it would totally change structure given that you have multiple meta_keys for single post + you can have duplicates! on post_id meta_key combination). meta_id makes sense here, but combination of post+meta_key index as well.

[galbaras]

Replying to johnbillion:

Additional indexes also take up disk space. Ironically the sites that would benefit the most from this additional index would also experience the greatest increase in space usage. (I'm not against the change, just pointing this out.)

In this day and age, disk space is cheap, but performance is golden.

Replying to fliespl:

Have you tested post_id, meta_key against meta_key, post_id?

Have you tested making either of those the primary index for the table?

It would be great if you could do this and report your findings.

[kkmuffme]

Any thoughts on making post_id the primary index (easier) or changing the way WordPress retrieves meta data (faster)?

This is impossible, since post id is not unique in meta.

Have you tested post_id, meta_key against meta_key, post_id?

The latter is faster, since generally you don't query by multiple post ids with a range or LIKE.

---

Looking at WooCommerce HPOS I see that they even use the meta value in an index, which does make sense, e.g. for range queries.

I think the existing "meta_key" only index can be dropped, since that's barely used.
I guess a meta_key, meta_value index like they use in WooCommerce HPOS would make more sense.
In addition to the meta_key, post_id index, potentially making this a 3 index as meta_key, meta_value, post_id

Since we've been using the initial index mentioned above since years, I don't have any benchmarks anymore and currently no time to get this further. Once we do performance analysis again in a few months and we still see room for improvement or a bottleneck, I'll provide a benchmark and a PR maybe via upgrade.php.

[galbaras]

I'm really hoping the core team (is there a performance team, or maybe the hosting team) would seriously consider this change. The imagify query above isn't typical, but running the usual WordPress testing through both the post_id, meta_key and meta_key, post_id options should be fairly easy and can truly improve WordPress performance.

Which index performs better is based on the cardinality of each column, and on most sites, post_id has the larger cardinality (1,000 posts with 50 meta keys is more likely than the reverse), but actually testing is best.

[pbearne]

Do we need we need to do this

Happy to create the patch