Sql-server – Inefficient queries on partitioned tables

Assuming that you have the primary key on a clustered index then the partitioning key needs to be part of the primary key.

You will not loose the benefit of partitioning by joining to non-partitioned tables, providing that the queries are designed to make use of the partitioned table, for example the following query WILL benefit from partitioning

SELECT F.Col1, F.Col2, D.Col3
FROM Fact_Partitioned F
    INNER JOIN Dim_MyDim D ON F.Col1 = D.Col1
WHERE F.Col1 = 5

But the following query WILL NOT benefit from partition elimination

SELECT F.Col1, F.Col2, D.Col3
FROM Fact_Partitioned F
    INNER JOIN Dim_MyDim D ON F.Col1 = D.Col1
WHERE D.Col1 = 5

It is a subtle difference, but in the first query, the join key is filtered in the partitioned table, taking advantage of elimination and then joined to the dimension. In the second query, the key is filtered in the dimension and then joined against the whole of the fact table, rather than just required partitions.

It goes without saying that the partitioning key needs to be in the WHERE clause for elimination to work, otherwise SQL Server does not know which partition(s) the data is in.

Adding a filter criteria on the JOIN clause will not help you. It needs to be in the WHERE clause to benefit from elimination.

The Partition Key does not need to be part of a non-clustered index (NCI) but if the NCI is unique, then it needs to contain the partitioning key in order to align the index. This is where the NCI is built on the same partition scheme as the table. NCIs should also be partition aligned unless there is an exceedingly good reason not to. I have never come across a good enough reason!

Expanding upon your list, here are a couple potential downsides that we have come across in real production workloads:

Seeking into multiple partitions

Expanding on the queries that do not use partition elimination could take longer to execute point, there is a specific pattern that is particularly affected: singleton seeks. This operation will become much slower if all (or even a modest subset of) partitions need to be accessed. The skip scan operation essentially performs a seek into every partition that cannot be eliminated.

Let's say that you have a billion row table (N = 1,000,000,000) with rows divided equally into 1,000 partitions (P = 1,000). A single seek is roughly O(log(N)) ~ 30 in a non-partitioned table. However, this same seek operation becomes roughly O(P*log(N/P)) ~ 20,000 in this hypothetical partitioned table. So the seek now performs over 500x more work if data from all partitions is needed (or sometimes even if it isn't needed, but SQL can't prove that based on your query).

Note that this can come up both when you explicitly query the table for one row (or a small range of rows) and in more complex queries when the partitioned table appears in the innder side of a loop join. The good news is that SQL Server is reasonably good about taking this into account in cost-based optimization, but that still typically means that you get a hash join when a loop-seek into a non-partitioned table would have been far more optimal.

Thread skew in parallel query execution

In parallel query plans, threads are allocated to partitions. If there is one partition that is much larger than the others, queries against the table may be particularly susceptible to thread skew. It's possible that one thread gets too high a proportion of rows and is processing long after the other threads have done their work. This situation can happen with non-partitioned tables as well, but any partition functions that do not equally distribute rows are particularly vulnerable.

See Parallel Query Execution Strategy for Partitioned Objects for a more detailed description of the allocation of threads to partitions. For example:

The query processor uses a parallel execution strategy for queries that select from partitioned objects. As part of the execution strategy, the query processor determines the table partitions required for the query and the proportion of threads to allocate to each partition. In most cases, the query processor allocates an equal or almost equal number of threads to each partition, and then executes the query in parallel across the partitions.