You say that the hierarchy gets modified. Presumably while this operation is running, there is some amount of blocking which is taking place then?
Even if the hierarchy is changing, are the roots for items changing?
Have you looked at the time it would take to just make the mapping table from root to item and index it?
I would like to see the execution plan to see what is happening - the CTE should get spooled, but as a manually materialized and indexed table it might perform better in the later steps.
Even with heavy activity, it would seem to me that someone has to be blocked if DML operations are changing data which this process is reading.
So I'd strongly consider taking a snapshot of the hierarchy.
In addition, you have a number of other INNER JOINs - you should review whether it is, in fact, the CTEs at all and whether there are any indexes which are missing to make those joins effective. The execution plan should tell you that.
You appear to have quite a few things in the WHERE clause which might help reduce some operations (and determine which indexes might be the best)), but it's hard to tell without looking at the execution plan or the indexes.
First off, you do not want to use char(50)
. Use varchar(50)
or just text
. Read more:
Assuming the following rules:
- Basic slugs never end with a dash.
- Duplicate slugs are suffixed with a dash and a sequential number (
-123
).
Note that all of the following methods are subject to a race conditions: concurrent operations might identify the same "free" name for the next slug.
To defend against it, you can impose a UNIQUE constraint on slug
and be prepared to repeat an INSERT upon duplicate key violation or you to take out a write lock on the table at the start of the transaction.
If you glue the suffix to the basic slug name with a dash and allow basic slugs to end in separate numbers, the specification is a tiny bit ambiguous (see comments). I suggest a unique delimiter of your choice instead (which is otherwise disallowed).
Efficient rCTE
WITH RECURSIVE
input AS (SELECT 'news-on-apple'::text AS slug) -- input basic slug here once
, cte AS (
SELECT slug || '-' AS slug -- append '-' once, if basic slug exists
, 1 as suffix -- start with suffix 1
FROM article
JOIN input USING (slug)
UNION ALL
SELECT c.slug, c.suffix + 1 -- increment by 1 ...
FROM cte c
JOIN article a ON a.slug = c.slug || c.suffix -- ... if slug-n already exists
)
(
SELECT slug || suffix AS slug
FROM cte
ORDER BY suffix DESC -- pick the last (free) one
LIMIT 1
) -- parentheses required
UNION ALL -- if the basic slug wasn't taken, fall back to that
SELECT slug FROM input
LIMIT 1;
Better performance without rCTE
If you worry about thousands of slugs competing for the same slug or generally want to optimize performance, I'd consider a different, faster approach.
WITH input AS (SELECT 'news-on-apple'::text AS slug
, 'news-on-apple-'::text AS slug1) -- input basic slug here
SELECT i.slug
FROM input i
LEFT JOIN article a USING (slug)
WHERE a.slug IS NULL -- doesn't exist yet.
UNION ALL
( -- parentheses required
SELECT i.slug1 || COALESCE(right(a.slug, length(i.slug1) * -1)::int + 1, 1)
FROM input i
LEFT JOIN article a ON a.slug LIKE (i.slug1 || '%') -- match up to last "-"
AND right(a.slug, length(i.slug1) * -1) ~ '^\d+$' -- suffix numbers only
ORDER BY right(a.slug, length(i.slug1) * -1)::int DESC
)
LIMIT 1;
If the basic slug isn't taken yet, the more expensive second SELECT
is never executed - same as above, but much more important here. Check with EXPLAIN ANALYZE
, Postgres is smart that way with LIMIT
queries. Related:
Check for the leading string and the suffix separately, so the LIKE
expression can use a basic btree index with text_pattern_ops
like
CREATE INDEX article_slug_idx ON article (slug text_pattern_ops);
Detailed explanation:
Convert the suffix to integer before you apply max()
. Numbers in text representation don't work.
Optimize performance
To get the optimum, consider storing the suffix separated from the basic slug and concatenate the slug as needed: concat_ws('-' , slug, suffix::text) AS slug
CREATE TABLE article (
article_id serial PRIMARY KEY
, title text NOT NULL
, slug text NOT NULL
, suffix int
);
The query for a new slug then becomes:
SELECT slug
|| COALESCE((
SELECT '-'::text || (max(suffix) + 1)::text
FROM article a
WHERE a.slug = i.slug), '') As slug
FROM (SELECT 'news-on-apple'::text AS slug) i -- input basic slug here
Ideally supported with a unique index on (slug, suffix)
.
Query for list of slugs
In any version of Postgres you can provide rows in a VALUES
expression.
SELECT *
FROM article
JOIN (
VALUES
('slug-foo'::text, 1)
('slug-bar',7)
) u(slug,suffix) USING (slug,suffix);
You can also use IN
with a set of row-type expressions Which is shorter:
SELECT *
FROM article
WHERE (slug,suffix) IN (('slug-foo', 1), ('slug-bar',7));
Details under this related question (as commented below):
For long lists, the JOIN
to a VALUES
expression is typically faster.
In Postgres 9.4 (released today!) you can also use the new variant of unnest()
to unnest multiple arrays in parallel.
Given an array of basic slugs and a corresponding array of suffixes (as per comment):
SELECT *
FROM article
JOIN unnest('{slug-foo,slug-bar}'::text[]
, '{1,7}'::int[]) AS u(slug,suffix) USING (slug,suffix);
Best Answer
This would produce the expected result: