I'm working on scripts that apply database schema updates. I've setup all my SQL update scripts using start transaction/commit. I pass these scripts to psql on the command line.
I now need to apply multiple scripts at the same time, and in one transaction. So far the only solution I've come up with is to remove the start transaction/commit from the original set of scripts, then jam them together inside a new start transaction/commit block. I'm writing perl scripts to do this on the fly.
Effectively I want nested transactions, which I can't figure out how to do in postgresql.
Is there any way to do or simulate nested transactions for this purpose? I have things setup to automatically bail out on any error, so I don't need to continue in the top level transaction if any of the lower ones fail.
Well you have the possibility to use nested transactions inside postgresql using SavePoints.
Take this code example:
CREATE TABLE t1 (a integer PRIMARY KEY);
CREATE FUNCTION test_exception() RETURNS boolean LANGUAGE plpgsql AS
$$BEGIN
INSERT INTO t1 (a) VALUES (1);
INSERT INTO t1 (a) VALUES (2);
INSERT INTO t1 (a) VALUES (1);
INSERT INTO t1 (a) VALUES (3);
RETURN TRUE;
EXCEPTION
WHEN integrity_constraint_violation THEN
RAISE NOTICE 'Rollback to savepoint';
RETURN FALSE;
END;$$;
BEGIN;
SELECT test_exception();
NOTICE: Rollback to savepoint
test_exception
----------------
f
(1 row)
COMMIT;
SELECT count(*) FROM t1;
count
-------
0
(1 row)
Maybe this will help you out a little bit.
I've ended up 'solving' my problem out of band - I use a perl script to re-work the input scripts to eliminate their start transaction/commit calls, then push them all into one file, which gets it's own start transaction/commit.
Related
I have two ETL jobs running on a stream I've created on a table. I need to run both on the same stream data and I read that in order to do so the DML statements (in my case merge statements) need to be wrapped in a transaction and committed at the end. I can't seem to be able to do that in a task though. I think I'm messing the semi-colon somewhere. This is what I've tried
create or replace task my_task as
begin;
merge into my_table1 t using my_stream s on t.id=s.id when matched insert values (id, col1);
merge into my_table2 t using my_stream s on t.id=s.id when matched insert values (id, col2);
commit;
This is just an example, the merge statements do more complex stuff.
The script just runs up to begin if I use a semi-colon or get an EOF error if I don't use one even though I have multiple semi-colons later in the script (so it tries to read past commit)
The task can call a stored procedure that contains the different statements within a transaction:
create procedure ...
as
$$
...
statement1;
BEGIN TRANSACTION;
statement2;
COMMIT;
statement3;
...
$$;
https://docs.snowflake.com/en/sql-reference/transactions.html
I am creating a function that will create a new table and insert informations about that table into other tables.
To create that table I am using the
CREATE TABLE IF NOT EXISTS
statement. Sadly it does not update the FOUND special variable in PostgreSQL nor can i find any other variable that would be updated.
Is there any way in PL/PGSQL to know whether that statement created a table or not?
The target of it is to not to double the informations in the other tables.
You may use CREATE TABLE AS in combination with ON_ERROR_ROLLBACK:
BEGIN;
-- Do inital stuff
\set ON_ERROR_ROLLBACK on
CREATE TABLE my_table AS
SELECT id, name FROM (VALUES (1, 'Bob'), (2, 'Mary')) v(id, name);
\set ON_ERROR_ROLLBACK off
-- Do remaining stuff
END;
To put it bluntly, with \set ON_ERROR_ROLLBACK on postgres will create a savepoint before each statement and automatically rollback to this savepoint or releasing it depending on the success of that statement.
The code above will execute initial and remaining stuff even if the table creation fails.
No, there are not any information if this command created table or not. The found variable is updated after query execution - not after DDL command. There is guaranteed so after this command, the table will be or this command fails to an exception.
Is there any way to write an SQL input file for sqlite that would somehow "throw" an error, eg. exited the transaction with rollback, if a condition isn't met?
I have a script that is supposed to do something, but only if there is a certain row in one table. If it's not there, the execution of the script might have fatal results and corrupt the db.
The script is only started on demand right now, but I would prefer to add a fail-safe which would prevent its execution in case there is some issue.
Basically what I need is something like
/* IF */ SELECT value FROM meta WHERE key = 'version' /* != hardcoded_version_string THROW SOME EXCEPTION */
Is there any way to accomplish that? In Postgre / Oracle this could be done using PLSQL but I am not sure if sqlite support any such a thing?
Triggers can use the RAISE function to generate errors:
CREATE VIEW test AS SELECT NULL AS value;
CREATE TRIGGER test_insert
INSTEAD OF INSERT ON test
BEGIN
SELECT RAISE(FAIL, 'wrong value')
WHERE NEW.value != 'fixed_value';
END;
INSERT INTO test SELECT 'fixed_value';
INSERT INTO test SELECT 'whatever';
Error: wrong value
Is there any way to write an SQL input file for sqlite that would
somehow "throw" an error, eg. exited the transaction with rollback, if
a condition isn't met?
One workaround may be to create dummy table and explicitly violate NULL constraint:
CREATE TABLE meta("key" VARCHAR(100));
INSERT INTO meta("key") VALUES ('version');
CREATE TEMPORARY TABLE dummy(col INT NOT NULL);
Transaction:
BEGIN TRANSACTION;
INSERT INTO dummy(col)
SELECT NULL -- explicit insert of NULL
FROM meta
WHERE "key" = 'version';
-- Error: NOT NULL constraint failed: dummy.col
-- rest code
INSERT INTO meta("key")
VALUES ('val1');
INSERT INTO meta("key")
VALUES ('val2');
-- ...
COMMIT;
SqlFiddleDemo
Keep in mind that SQLite is not procedural language and this solution is a bit ugly.
I've got a nightly archive job that sweeps existing rows from a records table into records_archive:
INSERT INTO records_archive
SELECT * FROM records WHERE id > 555 AND id <= 556;
For reasons that are outside the scope of this question, there is a column -- master_guid -- in both records and records_archive that has a UNIQUE index, and I can't change that. Rebuilding the index into a non-unique one is off the table for reasons out of my control. So, in principle, every master_guid value is supposed to be unique. There are some bad client implementations out there that sometimes fail to generate unique enough GUIDs, though, which results in a collision when we attempt to insert a record into records_archive that with a master_guid value that already exists in records_archive.
I can't fix the clients, so I need to work around it. The way to do that is to catch the unique_violation exception, modify the GUID (adding some random characters to it), and attempt to re-insert.
I can't just wrap the above-mentioned INSERT query in a stored procedure and catch the unique_violation exception, because the whole query is one transaction. I need row-level access. So, I wrote a stored procedure to iterate over each row and catch a unique_violation exception for that row individually:
CREATE OR REPLACE FUNCTION archive_records(start_id bigint,
end_id bigint)
RETURNS void
AS $$
DECLARE
_r record;
BEGIN
FOR _r IN
SELECT * FROM records WHERE id > start_id AND id <= end_id
LOOP
BEGIN
INSERT INTO records_archive VALUES (_r.*);
EXCEPTION WHEN unique_violation THEN
-- Manipulate the _r.master_guid value, add some random
-- numbers to it or whatever, and attempt reinsertion.
END;
END LOOP;
END;
$$ LANGUAGE 'plpgsql';
The problem is that records (and records_archive) are insanely wide tables. I don't want to explicitly enumerate every single column to be copied from _r to records_archive, not only because I'm lazy, but because this stored procedure would become a dependency in any future column changes in those tables.
The problem I've got is that this doesn't work, syntactically or conceptually:
INSERT INTO records_archive VALUES (_r.*);
Neither do any of these:
INSERT INTO records_archive _r;
INSERT INTO records_archive _r.*;
Is there a way to pull this off? If not, is there a better way to accomplish what I'm trying to accomplish?
I would create a BEFORE INSERT Trigger on records_archive. Do a select on the records_archive for the NEW.master_guid, and if it exists, manipulate it to add your random numbers.
You'd probably want a loop around the check to ensure the modified master_guid still didn't exist before went ahead with the insertion.
I wrote a function to create posts for a simple blogging engine:
CREATE FUNCTION CreatePost(VARCHAR, TEXT, VARCHAR[])
RETURNS INTEGER AS $$
DECLARE
InsertedPostId INTEGER;
TagName VARCHAR;
BEGIN
INSERT INTO Posts (Title, Body)
VALUES ($1, $2)
RETURNING Id INTO InsertedPostId;
FOREACH TagName IN ARRAY $3 LOOP
DECLARE
InsertedTagId INTEGER;
BEGIN
-- I am concerned about this part.
BEGIN
INSERT INTO Tags (Name)
VALUES (TagName)
RETURNING Id INTO InsertedTagId;
EXCEPTION WHEN UNIQUE_VIOLATION THEN
SELECT INTO InsertedTagId Id
FROM Tags
WHERE Name = TagName
FETCH FIRST ROW ONLY;
END;
INSERT INTO Taggings (PostId, TagId)
VALUES (InsertedPostId, InsertedTagId);
END;
END LOOP;
RETURN InsertedPostId;
END;
$$ LANGUAGE 'plpgsql';
Is this prone to race conditions when multiple users delete tags and create posts at the same time?
Specifically, do transactions (and thus functions) prevent such race conditions from happening?
I'm using PostgreSQL 9.2.3.
It's the recurring problem of SELECT or INSERT under possible concurrent write load, related to (but different from) UPSERT (which is INSERT or UPDATE).
This PL/pgSQL function uses UPSERT (INSERT ... ON CONFLICT ..) to INSERT or SELECT a single row:
CREATE OR REPLACE FUNCTION f_tag_id(_tag text, OUT _tag_id int)
LANGUAGE plpgsql AS
$func$
BEGIN
SELECT tag_id -- only if row existed before
FROM tag
WHERE tag = _tag
INTO _tag_id;
IF NOT FOUND THEN
INSERT INTO tag AS t (tag)
VALUES (_tag)
ON CONFLICT (tag) DO NOTHING
RETURNING t.tag_id
INTO _tag_id;
END IF;
END
$func$;
There is still a tiny window for a race condition. To make absolutely sure we get an ID:
CREATE OR REPLACE FUNCTION f_tag_id(_tag text, OUT _tag_id int)
LANGUAGE plpgsql AS
$func$
BEGIN
LOOP
SELECT tag_id
FROM tag
WHERE tag = _tag
INTO _tag_id;
EXIT WHEN FOUND;
INSERT INTO tag AS t (tag)
VALUES (_tag)
ON CONFLICT (tag) DO NOTHING
RETURNING t.tag_id
INTO _tag_id;
EXIT WHEN FOUND;
END LOOP;
END
$func$;
db<>fiddle here
This keeps looping until either INSERT or SELECT succeeds.
Call:
SELECT f_tag_id('possibly_new_tag');
If subsequent commands in the same transaction rely on the existence of the row and it is actually possible that other transactions update or delete it concurrently, you can lock an existing row in the SELECT statement with FOR SHARE.
If the row gets inserted instead, it is locked (or not visible for other transactions) until the end of the transaction anyway.
Start with the common case (INSERT vs SELECT) to make it faster.
Related:
Get Id from a conditional INSERT
How to include excluded rows in RETURNING from INSERT ... ON CONFLICT
Related (pure SQL) solution to INSERT or SELECT multiple rows (a set) at once:
How to use RETURNING with ON CONFLICT in PostgreSQL?
What's wrong with this pure SQL solution?
CREATE OR REPLACE FUNCTION f_tag_id(_tag text, OUT _tag_id int)
LANGUAGE sql AS
$func$
WITH ins AS (
INSERT INTO tag AS t (tag)
VALUES (_tag)
ON CONFLICT (tag) DO NOTHING
RETURNING t.tag_id
)
SELECT tag_id FROM ins
UNION ALL
SELECT tag_id FROM tag WHERE tag = _tag
LIMIT 1;
$func$;
Not entirely wrong, but it fails to seal a loophole, like #FunctorSalad worked out. The function can come up with an empty result if a concurrent transaction tries to do the same at the same time. The manual:
All the statements are executed with the same snapshot
If a concurrent transaction inserts the same new tag a moment earlier, but hasn't committed, yet:
The UPSERT part comes up empty, after waiting for the concurrent transaction to finish. (If the concurrent transaction should roll back, it still inserts the new tag and returns a new ID.)
The SELECT part also comes up empty, because it's based on the same snapshot, where the new tag from the (yet uncommitted) concurrent transaction is not visible.
We get nothing. Not as intended. That's counter-intuitive to naive logic (and I got caught there), but that's how the MVCC model of Postgres works - has to work.
So do not use this if multiple transactions can try to insert the same tag at the same time. Or loop until you actually get a row. The loop will hardly ever be triggered in common work loads anyway.
Postgres 9.4 or older
Given this (slightly simplified) table:
CREATE table tag (
tag_id serial PRIMARY KEY
, tag text UNIQUE
);
An almost 100% secure function to insert new tag / select existing one, could look like this.
CREATE OR REPLACE FUNCTION f_tag_id(_tag text, OUT tag_id int)
LANGUAGE plpgsql AS
$func$
BEGIN
LOOP
BEGIN
WITH sel AS (SELECT t.tag_id FROM tag t WHERE t.tag = _tag FOR SHARE)
, ins AS (INSERT INTO tag(tag)
SELECT _tag
WHERE NOT EXISTS (SELECT 1 FROM sel) -- only if not found
RETURNING tag.tag_id) -- qualified so no conflict with param
SELECT sel.tag_id FROM sel
UNION ALL
SELECT ins.tag_id FROM ins
INTO tag_id;
EXCEPTION WHEN UNIQUE_VIOLATION THEN -- insert in concurrent session?
RAISE NOTICE 'It actually happened!'; -- hardly ever happens
END;
EXIT WHEN tag_id IS NOT NULL; -- else keep looping
END LOOP;
END
$func$;
db<>fiddle here
Old sqlfiddle
Why not 100%? Consider the notes in the manual for the related UPSERT example:
https://www.postgresql.org/docs/current/plpgsql-control-structures.html#PLPGSQL-UPSERT-EXAMPLE
Explanation
Try the SELECT first. This way you avoid the considerably more expensive exception handling 99.99% of the time.
Use a CTE to minimize the (already tiny) time slot for the race condition.
The time window between the SELECT and the INSERT within one query is super tiny. If you don't have heavy concurrent load, or if you can live with an exception once a year, you could just ignore the case and use the SQL statement, which is faster.
No need for FETCH FIRST ROW ONLY (= LIMIT 1). The tag name is obviously UNIQUE.
Remove FOR SHARE in my example if you don't usually have concurrent DELETE or UPDATE on the table tag. Costs a tiny bit of performance.
Never quote the language name: 'plpgsql'. plpgsql is an identifier. Quoting may cause problems and is only tolerated for backwards compatibility.
Don't use non-descriptive column names like id or name. When joining a couple of tables (which is what you do in a relational DB) you end up with multiple identical names and have to use aliases.
Built into your function
Using this function you could largely simplify your FOREACH LOOP to:
...
FOREACH TagName IN ARRAY $3
LOOP
INSERT INTO taggings (PostId, TagId)
VALUES (InsertedPostId, f_tag_id(TagName));
END LOOP;
...
Faster, though, as a single SQL statement with unnest():
INSERT INTO taggings (PostId, TagId)
SELECT InsertedPostId, f_tag_id(tag)
FROM unnest($3) tag;
Replaces the whole loop.
Alternative solution
This variant builds on the behavior of UNION ALL with a LIMIT clause: as soon as enough rows are found, the rest is never executed:
Way to try multiple SELECTs till a result is available?
Building on this, we can outsource the INSERT into a separate function. Only there we need exception handling. Just as safe as the first solution.
CREATE OR REPLACE FUNCTION f_insert_tag(_tag text, OUT tag_id int)
RETURNS int
LANGUAGE plpgsql AS
$func$
BEGIN
INSERT INTO tag(tag) VALUES (_tag) RETURNING tag.tag_id INTO tag_id;
EXCEPTION WHEN UNIQUE_VIOLATION THEN -- catch exception, NULL is returned
END
$func$;
Which is used in the main function:
CREATE OR REPLACE FUNCTION f_tag_id(_tag text, OUT _tag_id int)
LANGUAGE plpgsql AS
$func$
BEGIN
LOOP
SELECT tag_id FROM tag WHERE tag = _tag
UNION ALL
SELECT f_insert_tag(_tag) -- only executed if tag not found
LIMIT 1 -- not strictly necessary, just to be clear
INTO _tag_id;
EXIT WHEN _tag_id IS NOT NULL; -- else keep looping
END LOOP;
END
$func$;
This is a bit cheaper if most of the calls only need SELECT, because the more expensive block with INSERT containing the EXCEPTION clause is rarely entered. The query is also simpler.
FOR SHARE is not possible here (not allowed in UNION query).
LIMIT 1 would not be necessary (tested in pg 9.4). Postgres derives LIMIT 1 from INTO _tag_id and only executes until the first row is found.
There's still something to watch out for even when using the ON CONFLICT clause introduced in Postgres 9.5. Using the same function and example table as in #Erwin Brandstetter's answer, if we do:
Session 1: begin;
Session 2: begin;
Session 1: select f_tag_id('a');
f_tag_id
----------
11
(1 row)
Session 2: select f_tag_id('a');
[Session 2 blocks]
Session 1: commit;
[Session 2 returns:]
f_tag_id
----------
NULL
(1 row)
So f_tag_id returned NULL in session 2, which would be impossible in a single-threaded world!
If we raise the transaction isolation level to repeatable read (or the stronger serializable), session 2 throws ERROR: could not serialize access due to concurrent update instead. So no "impossible" results at least, but unfortunately we now need to be prepared to retry the transaction.
Edit: With repeatable read or serializable, if session 1 inserts tag a, then session 2 inserts b, then session 1 tries to insert b and session 2 tries to insert a, one session detects a deadlock:
ERROR: deadlock detected
DETAIL: Process 14377 waits for ShareLock on transaction 1795501; blocked by process 14363.
Process 14363 waits for ShareLock on transaction 1795503; blocked by process 14377.
HINT: See server log for query details.
CONTEXT: while inserting index tuple (0,3) in relation "tag"
SQL function "f_tag_id" statement 1
After the session that received the deadlock error rolls back, the other session continues. So I guess we should treat deadlock just like serialization_failure and retry, in a situation like this?
Alternatively, insert the tags in a consistent order, but this is not easy if they don't all get added in one place.
I think there is a slight chance that when the tag already existed it might be deleted by another transaction after your transaction has found it. Using a SELECT FOR UPDATE should solve that.