7

Context

I have stumbled on an issue in SQL Azure where the Sort-operator is spilling over into TempDB due to poor row estimates.

  • We are querying a master table that is joined with a number of detail tables
  • The table uses a TenantId-column to partition the table per tenant
  • Some tenants have 10,000s of rows, some only 100s.
  • There is a row-level security policy in place that adds a FILTER PREDICATE to all queries on the aforementioned TenantId
  • Queries are generated by Entity-Framework in a .NET application
  • All index statistics are up-to-date
  • All detail tables rows are retrieved via Index Seeks

The issue

The Cardinality Estimator is yielding very low estimates due to the large variance of rowcounts between tenants. This combined with two inner joins reducing the estimates even further, makes it so that a query that actually yields 3600 rows, was only expected to yield 3. That's 3 orders of magnitude off.

What have I tried?

  1. Defined Filtered Statistics for those key values that yield a high number of rows, as an extra hint to the CE.
  2. Ran into limitations in dealing with parameterized queries. The OPTION ( RECOMPILE ) works for some predicates, but not for the TenantId which is injected via the aforementioned security policy.
  3. Inlined the filter predicate so we are effectively filtering on the same column twice works, but seems ... redundant to say the least
  4. Changing the INNER JOINs to LEFT OUTER JOINs improves the bad join estimates but since we use Entity-Framework I would prefer a solution which does not require changing the query. Note: obviously if the only way is to rewrite the query, then that is the route we will go.

Other ideas

  • I have toyed with the idea to just add a dummy tenant with 100k records to offset the estimations so that row estimates are at least large enough for the largest real tenant, but this would make us over estimate for the small ones.

What am I looking for?

  • Am I doing something wrong - did I paint myself into a corner?
  • Are there alternatives I could consider?

I welcome any ideas you might have, Thanks!


The sort operator is there for paging. I don't actually want to retrieve all the rows. So in short, the sort needs to happen in the db (not in the app).

Also, to be clear, the issue here is not the EF generated query. It is a simple query with a number of INNER/LEFT OUTER joins and a few filter predicates. It is not the typical 5000-line SQL statement that is legible to no one.

It seems the details lookups are done via hash match which loses the index's sort ordering. Also the column that is ordered on is user defined, and changes depending on the use-case, so to avoid the sort I would need an index per column they could possibly sort on.

Additional information

The Row-level security policy

CREATE FUNCTION Security.myAccessPredicate(@TenantId nvarchar(128))
    RETURNS TABLE WITH SCHEMABINDING
    AS
    RETURN SELECT 1 AS accessResult
    WHERE @TenantId = CAST(SESSION_CONTEXT(N'TenantId') AS nvarchar(128))
    OR CAST(SESSION_CONTEXT(N'TenantId') AS nvarchar(128)) LIKE 'ReservedTenantIdForCrossTenantOperations'");

CREATE SECURITY POLICY Security.mySecurityPolicy
    ADD FILTER PREDICATE Security.myAccessPredicate(TenantId) ON Schema1.Object10,
    ADD BLOCK PREDICATE Security.myAccessPredicate(TenantId) ON Schema1.Object10,
    ... And all the other tables

The Master table index DDL

CREATE NONCLUSTERED INDEX [IX_Object10] ON Schema1.Object10
(
    [TenantId] ASC,
    [Discriminator] ASC,
    [IsDeleted] ASC,
    [CreatedAt] ASC
)
INCLUDE (
    [Id],
    [Column3],
    ...
    [Column37]) 
WITH (
    PAD_INDEX = OFF
    , STATISTICS_NORECOMPUTE = OFF
    , SORT_IN_TEMPDB = OFF
    , DROP_EXISTING = OFF
    , ONLINE = OFF
    , ALLOW_ROW_LOCKS = ON
    , ALLOW_PAGE_LOCKS = ON
)

The query (anonymized)

EXEC sp_set_session_context @key=N'TenantId', @value=N'TenantId123'

SELECT *
FROM ( SELECT 
    Object2.Column2 AS Column2, 
    Object2.Column3 AS CreatedAt, 
    -- Removed some extra columns from Object2
    Object3.Column4 AS Column3, 
    Object4.Column4 AS Column5, 
    Object5.Column4 AS Column6, 
    Object6.Column7 AS Column7, 
    Object7.Column10 AS Column10, 
    Object8.Column10 AS Column11, 
    Object9.Column4 AS Column20, 
    CASE WHEN (Variable1 = ?) THEN Object2.Column26 ELSE cast(? as decimal(18)) END AS Column21, 
    CASE WHEN (Variable2 = ?) THEN Object2.Column27 ELSE ? END AS Column22, 
    CASE WHEN (Variable3 = ?) THEN Object2.Column28 ELSE ? END AS Column23, 
    CASE WHEN (Variable4 = ?) THEN Object2.Column29 END AS Column24
    FROM Schema1.Object10 AS Object2
    INNER JOIN Schema1.Object11 AS Object3 ON Object2.Column30 = Object3.Column2
    INNER JOIN Schema1.Object12 AS Object4 ON Object2.Column31 = Object4.Column2
    LEFT OUTER JOIN Schema1.Object10 AS Object5 ON (Object2.Column32 = Object5.Column2) AND (Object5.Column33 = ?)
    LEFT OUTER JOIN Schema1.Object10 AS Object6 ON (Object2.Column34 = Object6.Column2) AND (Object6.Column33 = ?)
    LEFT OUTER JOIN Schema1.Object13 AS Object7 ON Object2.Column35 = Object7.Column2
    LEFT OUTER JOIN Schema1.Object13 AS Object8 ON Object2.Column36 = Object8.Column2
    LEFT OUTER JOIN Schema1.Object14 AS Object9 ON Object2.Column37 = Object9.Column2
    WHERE (Object2.Discriminator = N'SomeCategory') AND (0 = Object2.IsDeleted)
)  AS Object1
ORDER BY Object1.CreatedAt DESC
OFFSET ? ROWS FETCH NEXT ? ROWS ONLY

The query plan (anonymized) : https://pastebin.com/msjPQ6Vs

Estimated

enter image description here

Actuals

enter image description here

The master table (Schema1.Object10 AS Object2) returns 3600+ actual rows from the index seek, but CE estimates 381 rows. Then the two inner joins to Schema1.Object11 and Schema1.Object12 further reduce the estimate to 2.9 estimated number of rows. This makes no sense to me. The inner joins happen on non-null id columns with foreign key constraints, so it should be impossible to not find a match.

The sort operator tooltip:

enter image description here

Actual Timings, whereas the cumulative time before the sort operator was 16ms

enter image description here

  • 1
    How much does it actually spill? If you're selecting a lot of columns you can try the techniques I talk about here: groupby.org/conference-session-abstracts/… – Erik Darling Oct 4 '18 at 17:33
  • 1
    Also share the table and index DDL and partition function. Note generally that the inherent difficulty query plan management in a multi-tenant database is one of the reasons to favor the database-per-tenant model in Azure SQL, or at least database-per-group-of-similar-tenants. – David Browne - Microsoft Oct 4 '18 at 18:53
  • @DavidBrowne-Microsoft I do not actually use a partition function. I chose to describe what we're doing as "partitioning" the data, but in SQL Server terminology, that might not be what we're doing here. To be fair, it's not only the TenantId we're filtering on, also a second Discriminator column. The same variance in cardinality that exists between rows that belong to a tenant, exist in terms of rows that have a certain discriminator value. So even if had architected a database per tenant, we would still run into these kinds of issues. – Yannick Motton Oct 4 '18 at 23:45
  • so, a question - technically, to query across tenants you have to have some kind of partitioning set up, or am I wrong? you would be using elastic queries, and then you have either vertical partitioning or horizontal partitioning(sharding) configured. I think there may be some limitations that can skew stats depending on which way it's done. – rottengeek Oct 24 '18 at 19:32
  • Have you tried to rebuild all the statistic objects on the table with fullscan and skip the filters on the statistics ? – Spörri Oct 26 '18 at 10:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.