Partitioned view aggregation query not optimized

Question

Given the following two tables:

CREATE TABLE SalesLedger (
  Id int PRIMARY KEY IDENTITY,
  Date date NOT NULL,
  Total decimal(38,18),
  INDEX IX (Date, Total)
);

CREATE TABLE Purchases (
  Id int PRIMARY KEY IDENTITY,
  Date date NOT NULL,
  Total decimal(38,18),
  INDEX IX (Date, Total)
);

And the following view

CREATE VIEW ViewMetrics
AS

Select
  Date,
  'Sale' as Metric,
  Total as Value
From SalesLedger

UNION ALL

Select
  Date,
  'Purchase' as Metric,
  Total as Value
From Purchases;

The following query uses a Concatenation Sort pair:

Select SUM(Value) as Sales, Date
from ViewMetrics
Group By Date;

PasteThePlan

---

Whereas a minor rewrite gives the obviously more performant Merge Concatenation

SELECT SUM(Sales), Date
FROM (
  Select SUM(Value) as Sales, Date
  from ViewMetrics
  Group By Metric, Date
) t
GROUP BY Date;

PasteThePlan

---

The compiler can clearly see that the view is partitioned by Metric, as this query shows, no Sort is needed:

Select SUM(Value) as Sales, Date
from ViewMetrics
where Metric = 'Sale'
Group By Date;

---

The question is: why does the first query force a Sort, whereas the second can use a more efficient Merge Concatenation, given that the Metric column has no WHERE predicate in either case?

Shouldn't the compiler be able to see that a Merge would work given that the indexes are sorted on Date already, and the partitioning is on Metric? Or if it cannot see that, why does GROUP BY Metric, Date suddenly give it that ability?

db<>fiddle

Even stranger, as @MartinSmith has found, if there is no data then the compiler will use the better plan, although without an intermediate aggregation on Metric, Date. db<>fiddle On the other hand, the merge without partial aggregation is probably slower than sorting after partial aggregation anyway, because there are more rows to merge. Question is why can't it do both partial aggregation and merge at the same time by default?

I'm guessing there is some specific optimization for a partitioned view when aggregation includes the partitioning, because it uses a Concatenation in that case and when an ordering is required then it uses Merge Concatenation, see db<>fiddle. This then helps it when you want to further aggregate, as the data is now sorted already in the correct order. But if you don't do the intermediate aggregation it has no logic that would apply it.

Answer 1

The SQL Server optimizer has two main ways to push an aggregate down past a union all.

1. Global Pushdown

The first rule is GbAggBelowUniAll. It's a fairly straightforward transform that moves the aggregation onto each of the union inputs.

It can only do this safely if the union is disjoint—that is, if there's something that makes each input completely independent, and that factor appears in the GROUP BY clause.

It doesn't have to be a literal value like in your case, but it does have to be something the optimizer can recognize as completely separating the sets, like non-overlapping ranges. As part of the transformation, the constant part of the grouping clause is removed.

This rule is engaged with your rewrite because the Metric attribute is disjoint and present in the grouping specification.

An example using the Adventure Works sample database:

-- Helpful indexes
CREATE INDEX i ON Production.TransactionHistory 
    (Quantity, TransactionDate);
CREATE INDEX i ON Production.TransactionHistoryArchive 
    (Quantity, TransactionDate);

SELECT 
    U.Quantity, 
    MAX(U.TransactionDate)
FROM 
(
    -- The predicates on quantity make the two union inputs disjoint
    SELECT 
        TH.*
    FROM Production.TransactionHistory AS TH
    WHERE 
        TH.Quantity BETWEEN 50 AND 60

    UNION ALL

    SELECT
        THA.*
    FROM Production.TransactionHistoryArchive AS THA
    WHERE 
        THA.Quantity BETWEEN 10 AND 20
) AS U
-- Grouping by the disjoint element
GROUP BY 
    U.Quantity
ORDER BY 
    U.Quantity
--OPTION (FORCE ORDER)
;

With the FORCE ORDER hint uncommented, the optimizer is prevented from moving the aggregate around:

Without the hint, the top-level aggregate can be moved and copied to each union input:

2. Local Aggregation

The second transformation involves a couple of different rules.

First, GenLGAgg splits an aggregate into two parts, a global aggregate and a local aggregate. For example, a COUNT aggregate would be split into a local COUNT aggregate and a global SUM aggregate that adds all the local contributions together to arrive at the correct result.

The general idea is used in both serial and parallel plans. Sometimes, the local aggregate computes a subtotal local to its own thread, sometimes it performs a subset of the work below a join. In any case, the general idea is the same: do some part of the overall aggregation task as early as possible.

Like many optimizer explorations, GenLGAgg produces one or more alternatives that can be further explored by other rules. For example, the new local aggregates may be past joins or matched to an indexed view.

In your case, a rule named LocalAggBelowUniAll is employed to move the local aggregate below the UNION ALL. This is what was happening with the original query.

Importantly, the local aggregate is not quite a normal aggregate. It is performing only part of the calculation and may end up on one of many threads in a parallel plan. A global aggregate always runs on a single thread to ensure correct results.

In a parallel plan, a local aggregate may be physically implemented as a Hash Match Partial Aggregate. This operator only gets a small, fixed memory grant and never spills to tempdb. If it runs out of memory, it simply stops aggregating. Results will still be correct thanks to the global aggregate.

The caveats are not limited to this physical operator or parallel plans. In general, you should think of a local aggregate as being a bit different from the normal kind you write in SQL.

This is mostly a consequence of implementation details and the need to preserve the connection between the local and global aggregates. Like many exploration rules, LocalAggBelowUniAll represents a query rewrite you could perform yourself, but you should not expect it to behave exactly the same in all respects as the closest T-SQL representation. It should also be said that the optimizer has the advantage of being able to dynamically decide which rewrite to use based on current statistics and metadata. This is not generally true for a manual rewrite.

Anyway, one of the consequences of a local aggregate being a bit different is that it does not come with a uniqueness guarantee associated with its grouping keys. This is important in many instances but particularly so with your desired Merge Concatenation operator.

Merge Concatenation

As the plan operator's name suggests, Merge Concatenation is just a normal Merge Join operator running in a special mode. It requires input sorted on the 'join keys' though exactly which sort order is needed can be affected by the projection column list, global ordering requirement, and any uniqueness guarantees available (see reference below).

A normal aggregate that provides grouping key uniqueness guarantees can allow the Merge Concatenation to require less onerous sorting than is possible with input from a local aggregate.

For your original query, the optimizer did consider the Merge Concatenation alternative, but the local aggregate meant that sorts were required to satisfy the required input properties:

The extra Sorts and higher Merge Concatenation cost meant the optimizer chose the cheaper plan option with a Concatenation and a single Sort. As always, these choices are driven from the cost model.

Other Notes

With empty tables, the optimizer costs the Merge Concatenation option cheapest because it merges only one row from each input. The higher per-row cost of a Merge Concatenation does not offset the cost of the extra Sort needed in the Concatenation plan.

Your query should reject nulls, or the base tables should have an explicit NOT NULL on the Value attribute. This will simplify the final plan (and make any future indexed view matching easier). In particular, the Stream Aggregates will no longer have to compute a COUNT_BIG(Total) aggregate and the Compute Scalars will no longer be needed.

You should also use schema prefixes and avoid keywords as attribute names.

SELECT 
    TotalSales = SUM(T.Sales), 
    T.[Date]
FROM 
(
    SELECT 
        Sales = ISNULL(SUM(VM.[Value]), 0.0), 
        VM.[Date]
    FROM dbo.ViewMetrics AS VM
    WHERE
        VM.[Value] IS NOT NULL
    GROUP BY
        VM.Metric, 
        VM.[Date]
) AS T
GROUP BY 
    T.[Date];

Further Reading

• SQL Server Optimizer Bug with JOIN and GROUP BY
  covers local, partial, and global aggregation
• Avoiding Sorts with Merge Join Concatenation
  details for merge concatenation sorts

Both written by me.