Sql-server – Isolation level issue

concurrencyisolation-levelsql serverupsert

I need some help in choosing the correct isolation level on my transactions.

My situation is as follows. I have a table with a unique constraint over two columns. For the sake of the argument, lets say the table is called Animal and the columns are Name and Owner. So every owner can only have one pet with a certain name.

My problem is that I'm receiving errors during inserts because they are coming in at such a phase that I'm getting exceptions triggered by the unique constraint. Inserts are happening by stored proc. At the start of the SP I do a select from Animal where Owner=x and Name=y to determine if the animal exists, if so, select the Id and move on, else insert the new animal. The rest of the proc uses the Id to insert/update Animal details in other tables.

I know I can set the isolation level to Serializable, but I'm hesitant to do so because I know that it can cause deadlocks and since I'm having to do loads of inserts rapidly, this is a very real threat.

Any advise will be welcomed, thanks.

Best Answer

Answer originally left by Dan Guzman in comments:

The method I usually use is an explicit transaction plus a UPDLOCK, HOLDLOCK locking hint on the SELECT query. That will help avoid deadlocks and avoid the race condition, without resorting to serializable isolation for the entire transaction.

Also SERIALIZABLE won't really help as the initial select will get shared (S) range locks. So both sessions will get the S lock, and then, instead of a PK violation, you'll get a deadlock.

With the UPDLOCK in addition to HOLDLOCK, only one session can select the same key regardless of whether it exists (holding update key lock) or not (holding update key range lock). This will serialize access to the specified key in this code path.

This worked for the OP, who implemented it as follows:

@Id = SELECT *
  FROM ANIMAL WITH(UPDLOCK, HOLDLOCK)
  WHERE Owner = x AND Name = y
IF(@id <> NULL)
  UPDATE ANIMAL
ELSE
  INSERT ANIMAL

--DO OTHER STUFF

Clarify `ON CONFLICT DO UPDATE` behavior

Consider the manual here:

For each individual row proposed for insertion, either the insertion proceeds, or, if an arbiter constraint or index specified by conflict_target is violated, the alternative conflict_action is taken.

Bold emphasis mine. So you do not have to repeat predicates for columns included in the unique index in the WHERE clause to the UPDATE (the conflict_action):

INSERT INTO test_upsert AS tu
       (name   , status, test_field  , identifier, count) 
VALUES ('shaun', 1     , 'test value', 'ident'   , 1)
ON CONFLICT (name, status, test_field) DO UPDATE
SET count = tu.count + 1;
WHERE tu.name = 'shaun' AND tu.status = 1 AND tu.test_field = 'test value'

The unique violation already establishes what your added WHERE clause would enforce redundantly.

Clarify partial index

Add a WHERE clause to make it an actual partial index like you mentioned yourself (but with inverted logic):

CREATE UNIQUE INDEX test_upsert_partial_idx
ON public.test_upsert (name, status)
WHERE test_field IS NULL;  -- not: "is not null"

To use this partial index in your UPSERT you need a matching conflict_target like @ypercube demonstrates:

ON CONFLICT (name, status) WHERE test_field IS NULL

Now the above partial index is inferred. However, as the manual also notes:

[...] a non-partial unique index (a unique index without a predicate) will be inferred (and thus used by ON CONFLICT) if such an index satisfying every other criteria is available.

If you have an additional (or only) index on just (name, status) it will (also) be used. An index on (name, status, test_field) would explicitly not be inferred. This doesn't explain your problem, but may have added to the confusion while testing.

Solution

AIUI, none of the above solves your problem, yet. With the partial index, only special cases with matching NULL values would be caught. And other duplicate rows would either be inserted if you have no other matching unique indexes / constraints, or raise an exception if you do. I suppose that's not what you want. You write:

The composite key is made up of 20 columns, 10 of which can be nullable.

What exactly do you consider a duplicate? Postgres (according to the SQL standard) does not consider two NULL values to be equal. The manual:

In general, a unique constraint is violated if there is more than one row in the table where the values of all of the columns included in the constraint are equal. However, two null values are never considered equal in this comparison. That means even in the presence of a unique constraint it is possible to store duplicate rows that contain a null value in at least one of the constrained columns. This behavior conforms to the SQL standard, but we have heard that other SQL databases might not follow this rule. So be careful when developing applications that are intended to be portable.

Allow null in unique column

I assume you want NULL values in all 10 nullable columns to be considered equal. It is elegant & practical to cover a single nullable column with an additional partial index like demonstrated here:

PostgreSQL multi-column unique constraint and NULL values

But this gets out of hand quickly for more nullable columns. You'd need a partial index for every distinct combination of nullable columns. For just 2 of those that's 3 partial indexes for (a), (b) and (a,b). The number is growing exponentially with 2^n - 1. For your 10 nullable columns, to cover all possible combinations of NULL values, you'd already need 1023 partial indexes. No go.

The simple solution: replace NULL values and define involved columns NOT NULL, and everything would work just fine with a simple UNIQUE constraint.

If that's not an option I suggest an expression index with COALESCE to replace NULL in the index:

CREATE UNIQUE INDEX test_upsert_solution_idx
    ON test_upsert (name, status, COALESCE(test_field, ''));

The empty string ('') is an obvious candidate for character types, but you can use any legal value that either never appears or can be folded with NULL according to your definition of "unique".

Then use this statement:

INSERT INTO test_upsert as tu(name,status,test_field,identifier, count) 
VALUES ('shaun', 1, null        , 'ident', 11)  -- works with
     , ('bob'  , 2, 'test value', 'ident', 22)  -- and without NULL
ON     CONFLICT (name, status, COALESCE(test_field, '')) DO UPDATE  -- match expr. index
SET    count = COALESCE(tu.count + EXCLUDED.count, EXCLUDED.count, tu.count);

Like @ypercube I assume you actually want to add count to the existing count. Since the column can be NULL, adding NULL would set the column NULL. If you define count NOT NULL, you can simplify.

Another idea would be to just drop the conflict_target from the statement to cover all unique violations. Then you could define various unique indexes for a more sophisticated definition of what's supposed to be "unique". But that won't fly with ON CONFLICT DO UPDATE. The manual once more:

For ON CONFLICT DO NOTHING, it is optional to specify a conflict_target; when omitted, conflicts with all usable constraints (and unique indexes) are handled. For ON CONFLICT DO UPDATE, a conflict_target must be provided.

Best Answer

Related Solutions

Sql-server – Why are timestamps not always increasing with concurrent inserts

PostgreSQL UPSERT – Handling NULL Values Issue

Clarify ON CONFLICT DO UPDATE behavior

Clarify partial index

Solution

Related Question

Clarify `ON CONFLICT DO UPDATE` behavior