Natural Data Clustering: Why Nested Loops Win So Often May, 2008 ©2008 Dan Tow, All rights reserved [email protected] SingingSQL.

Natural Data Clustering:Why Nested Loops Win So Often

May, 2008

©2008 Dan Tow, All rights [email protected]

SingingSingingSQL SQL PresentsPresents:

Nested Loops vs. Hash Joins

Joining A to B, assuming A has a better filter than B (almost certainly true for a correct nested-loops join order):

Nested Loops: For each A, find matching Bs following the indexed join key.

Hash Join: Reach B as if with a single-table query, then line up A-to-B matches on the hash of the join keys.

Common Cases to Consider

•Nested-loops join reaches a large fraction of a small joined-to table: Optimizer tends to prefer a hash join, correctly, but both alternatives are well-cached and fast.

•Nested loops obviously reach a tiny fraction of a large joined-to table: Optimizer tends to prefer a nested-loops join, correctly, and this is much faster than the hash-join alternative.

Common Cases to Consider

•Nested loops reach a tiny fraction of a large joined-to table, but the optimizer greatly over-estimates that fraction: Optimizer would prefer a nested-loops join, if it knew more, and this is much faster than the hash-join alternative.

–This tends to happen for “exceptions” reports, where two conditions are almost mutually exclusive, but the optimizer’s assumption of statistically independent conditions yields the over-estimate.–As humans, we can and should expect reasonable rowcounts in reports, so we should often expect this case.

A Rare Case to Consider

•Nested loops reach a large fraction of a large joined-to table: Optimizer will strongly (and usually correctly) prefer a hash join to a full table scan, taking into account the benefit of multi-block reads, and avoiding multiple logical I/Os to the same blocks.

–When the join order is correct, this almost always implies a huge set of rows being returned by the query.–Such large return rowsets are very rarely useful for human consumption, but can reasonably occur in middleware or conversion processes.

Common Cases to Consider

•Nested loops reach a medium-sized fraction of a large joined-to table: Optimizer will mildly prefer a hash join to a full table scan, taking into account the benefit of multi-block reads, and avoiding multiple logical I/Os to the same blocks.

–In this common case, nested-loops are frequently preferable to the hash-join choice of the optimizer, largely owing to self-caching in queries of naturally clustered tables.

A Business Database, from 10,000 Meters High

•Tables that grow the largest track frequent business events, such as product orders, customer calls, account payments, et cetera.

•The application exists to drive correct business actions and choices, and these in turn depend on recent events.

•Therefore, most data queried from large tables concerns recent events, and was recently added to those tables, clustered at the top.

A Business Database, from 10,000 Meters High

•Event-type tables rarely stand alone – they usually join to several other events-related tables tracking related events and event details.

•Joins to large tables are therefore usually from recent event-type data in one table to related, recent, event-type data in another table.

Related, Event-Type Heap Tables

Recent Master rows

Recent Detail rows

A Well-Clustered Range Will Likely Join to a Well-Clustered Range!

Recent Master rows

Recent Detail rows

Nested-Loops Costs to Consider

•We can reach 30,000 rows or more per second with nested loops to cached data, a trivial time compared to the magnitude of the output, unless a bad join order reaches tables in the wrong order! (What application must produce multiple 30,000-plus-row reports per minute, even in the worst case?)•Where nested-loops costs matter, it is because physical I/O time is high, or the join order is wrong.

Nested-Loops Index Physical I/O

•If the join key is an ordinary sequence-generated ID, the join-key index will naturally cluster in the order the rows were created.

•In this case, the query need only cache a few blocks to read thousands of rows from the recent end of the table – index-block I/O costs are trivial thanks to self-caching early in the query execution, even if these blocks are not yet cached.

Nested-Loops Table Physical I/O

•In the ideal case (which is surprisingly close to the common case!), the joined-to table is accessed roughly in row-creation order.•Read-ahead in the disk subsystem effectively converts reads like this into multi-block reads.•These two effects lead to physical I/O much like a fractional table scan of only the top of the heap table (which is likely largely cached, anyway).

Physical Demonstration of Co-clustered Performance

•Two million-row tables, joined one-to-one, matching rows created in the same order, reading the most-recent 1% of the rows:

Join Type Time/s LIO PIO

Nested-Loops 0.74 30302 314

Hash 6.7 12972 1751

Physical Demonstration of Co-clustered Performance

•Two million-row tables, joined one-to-one, matching rows created in the same order, reading the most-recent 1% of the rows:

Join Type Time/s LIO PIO CBO “Cost”

Nested-Loops 0.74 30302 314 20337

Hash 6.7 12972 1751 3737

Physical Demonstration of Non-Co-clustered Performance

•Two million-row tables, joined one-to-one, joined-to matching rows created in scattered order, reading the most-recent 1% of the driving-table rows:

Join Type Time/s LIO PIO

Nested-Loops 49.2 30223 10029

Hash 6.7 12899 1592

Physical Demonstration of Non-Co-clustered Performance

•Two million-row tables, joined one-to-one, joined-to matching rows created in random order, reading the most-recent 1% of the driving-table rows:

Join Type Time/s LIO PIO CBO “Cost”

Nested-Loops 49.2 30223 10029 20337

Hash 6.7 12899 1592 3737

Comparing Co-Clustered and Non-Co-clustered Performance

Join Type Time/s LIO PIO CBO “Cost”

Co-clustered Nested Loops

0.74 30302 314 20337

Non-co-clustered NL

49.2 30223 10029 20337

Hash (both) 6.7 13K 1.6K-1.8K

3737

Comparing Co-Clustered and Non-Co-clustered Performance

Join Type Time/s LIO PIO CBO “Cost”

Co-clustered Nested Loops

0.74 30302 314 20337

Non-co-clustered NL

49.2 30223 10029 20337

Hash (both) 6.7 13K 1.6K-1.8K

3737

•Cost function predicts relative runtimes well in non-co-clustered case!•Co-clustered case is more important, though, and greatly favors nested loops!

Conclusions

•The join method to small tables matters little, but tends to slightly favor hash joins for queries returning substantial rowcounts.•Hash joins to large tables can win in rare cases where sensible queries return (or summarize) very high rowcounts.

Conclusions

•Hash joins to large tables can win in cases where the join order is wrong, or where join-key indexes are missing or disabled through join-key type conversions or other index-disabling functions.•(In the above case, the better solution is to enable the optimum nested-loops plan!)

Conclusions

•Nested loops almost always win when the CBO prefers them.•Hint-forced nested loops usually win when the CBO underestimates the combined selectivity of multiple filters (when mutually-exclusive conditions violate the assumption of statistically independent conditions).

Conclusions•Hint-Forced nested loops in the best join order usually win in the common case of joins from recent rows to related recent rows (joins between naturally co-clustered tables).•Hash joins to large full table scans are rarely better than the best nested-loops plan.

Questions?