Comprehensive Indexing via Automated Execution Plan Analysis (ExecStats) Joe Chang jchang6 @ yahoo www.qdpma.com Slide deck here

About Joe • • • •

SQL Server consultant since 1999 Query Optimizer execution plan cost formulas (2002) True cost structure of SQL plan operations (2003?) Database with distribution statistics only, – no data 2004

• Decoding statblob-stats_stream – writing your own statistics

• Disk IO cost structure • Tools for system monitoring, execution plan analysis See ExecStats http://www.qdpma.com/ExecStats/SQLExecStats.html Download: http://www.qdpma.com/ExecStatsZip.html Blog: http://sqlblog.com/blogs/joe_chang/default.aspx

Objectives • Complexities & depth SQL performance – Cause and Effect

• Focus on the execution plan – Inefficient plans – missing indexes – very large estimate/actual row discrepancies

• Comprehensive Index Strategy – few good indexes, but no more than necessary

Non-Objective • List of rules to be followed blindly • without consideration for the underlying reason

• and whether rule actually applies in the current circumstance DBA skill: cause and effect analysis & assessment

Preliminary: Correct Results • Normalization – Data stored once, avoid anomalies

• Unique Keys – Avoid duplicate rows

• Foreign Keys – Avoid orphaned rows Incorrect architecture requires use of SELECT DISTINCT etc. to correct architecture deficiencies Which may cause performance problems as well Correct action is to address the architecture mistakes before the performance issue.

Performance Big Picture SQL

Tables natural keys

Indexes

Query Optimizer DOP Memory Parallel plans API Server Cursors: open, prepare, execute, close?

SET NO COUNT Information messages

Execution Plan Storage Engine Hardware

Statistics & Compile parameters

Compile Row estimate propagation errors

Recompile temp table / table variable

Index & Stats Maintenance

Tables and SQL combined implement business logic Natural keys with unique indexes, not SQL Index and Statistics maintenance policy 1 Logic may need more than one execution plan? Compile cost versus execution cost? Plan cache bloat?

The Execution Plan links all the elements of performance Index tuning alone has limited value Over indexing can cause problems as well

Indexing Principles • Good cluster key choice – Grouping + unique, not too wide

• Good nonclustered indexes – For key queries, not necessarily every query – Covered indexes where practical – Create and drop custom indexes for maintenance ops/special circum.

• No more indexes than necessary – Update overhead – Compile overhead – May tolerate occasional scans to avoid update maintenance

Note emphasis on good, not perfect

SQL – Plan – Index Usage dm_exec_query_stats

Execution Plan

dm_db_index_usage_stats

Using DMVs – Execution Plan dm_exec_query_stats

dm_exec_query_plan dm_exec_text_query_plan

dm_exec_sql_text System views

Execution Plan Indexes, joins Compile parameters

dm_db_index_usage_stats dm_db_index_operational_stats dm_db_index_physical_stats

dm_exec_query_profiles 2014 Real time query progress?

Indexes, key columns, Include list, filter, XML, Columns store etc.

DBCC SHOW_STATISTICS STATS_DATE(object_id, stats_id)

dm_db_stats_properties sys.dm_db_stats_properties, is available in SQL Server 2012 starting with Service Pack 1 and in SQL Server 2008 R2 starting with SP2. last_updated, rows, rows_sampled, steps, unfiltered_rows, modification_counter

Performance Analysis • Getting Top SQL from dm_exec_query_stats – Manually examining top execution plans

• Index Reduction – dm_db_index_usage_stats – Drop unused indexes (based on long period) – Consolidating indexes with similar keys – Infrequently used indexes? • Must hunt down SQL, possibly low item in query stats • Can it use another index?

Real World Example

July 2014 3.6B rows, 3.4TB, 1.5TB data 1.8TB indexes Key tables have 21, 8, 14, 4 and 14 nonclustered indexes

Largest Table – 21 (NC) Indexes

Index Reduction

Dec 2014

4.6B rows, 2.3TB, 1.7TB data 0.5TB indexes Key tables have 6, 4, 3, 2 and 4 nonclustered indexes

Nonclustered indexes reduced from 21 to 6

Index Reduction + Compression

Feb 2015

4.9B rows, 0.9TB,

0.55TB data

0.34TB indexes

Key tables have 7, 5, 3, 2 and 4 nonclustered indexes

1 index added, 1 index awaiting removal, possibly 2

Compression Notes • Very high compression was achieved – Because all keys were 16-byte GUID – Even on dimensions, when natural key would have been 1, 2 or 4 bytes!

• Core data + indexes – 648GB data 230GB indexes w/o compression – 174GB data 185GB indexes w/compression

• Reduction in I/O even with SSD storage far outweigh compression overhead! System memory: 256GB (220)

Storage: Violin (NAND Flash)

Index reduction Compression & statistics

HDD Storage

Violin GDC

Database view

File IO view

Table view columns

Indexes - continued

Literal identifying the execution plans Number of execution plans that reference the index in Seeks, Scans, that reference the index in Seeks, Lookups, Insert/Updates and Deletes Scans, Lookups, Insert/Upd & Deletes

Query Execution Stats - 1

Query Execution Stats - more

Dataspace – Partition Scheme view

Partition View

Procedure and Functions • • • • • •

Columns Dbid, schema, object, object_id, type, Create date, modify date, Number of references (NumRef) (literal) plan reference (from QExec Stats) Caller reference (Functions only)

Volumes

Slides not used

Performance Strategy • Tables – support business logic – Normalization, uniqueness etc.

Identity key / alternative: large & small customers

• SQL – clear SARG, Query optimizer interpretable – 1 Logic maps to X Execution plans

• Indexes – good cluster key choice – Good nonclustered indexes, no more than necessary

• Statistics – sample strategy & update frequency • Compile parameter strategy • •

Temp table / Table variable strategy: Recompile & Row est. prop. error Parallel execution plans: DOP and CTOP strategy

• Identify (weight) important SQL statements – stored procedure: parameter values & code path • Recompile impact for temp tables

• Execution plan cross references SQL & indexes – Actual plan is better than estimate plan – Compile parameters & skewed statistics • Temp tables - Recompile impact

Automate Execution Plan analysis to fully cross-reference SQL to index usage

SQL & Execution Plan Sources • Estimated Execution Plan – dm_exec_query_stats • Contents of plan cache + execution statistics

– List of stored procedures • SELECT name FROM sys.procedures

• Any SQL list – Plans not in cache, to be generated – Can also execute SQL for actual plans

sys.dm_exec_query_stats • sql_handle – token for batch or stored procedure

• statement_start_offset – sql_handle + offset = SQL statement

• plan_handle – SQL (batch) can have multiple plans on recompile

• query_hash – identify queries with similar logic, – differing only by literal values

sys.procedures • Get list of stored procedures in database – functions are called from procedure?

• Generate estimated execution plan for each – Default parameters

• Full map of index usage to stored procedure • No trigger details in estimated plan

SQL List • Configuration file has SQL to retrieve SQL list – Can be • explicit SQL • or stored procedures with parameters

– Same procedure, multiple parameter set • To expose different code path (actual plan) • EXEC proc WITH @P1 RECOMPILE (estimated plan)

About ExecStats •

General information

• Execution plan sources 1. 2. 3. 4.

• • • • •

dm_exec_query_stats list of all stored procedures (estimated) List of SQL in table (estimated or actual plan) Trace file

Correlates execution plans to index usage Procedures, functions and triggers Rollup file IO stats by DB, filegroup, disk/vol, data/log Distribution Statistics Output to Excel, sqlplan file, (sql in txt file)

ExecStats Output Files • • • • •

Txt – runtime info Log – abbreviated SQL error logs Excel – Missing Indexes DMV SQL plan directory

This can be sent to someone who can identify and fix your problem

Important Items • Query cost – plan efficiency? Recompiles? – Compile parameters – skewed statistics

• CPU versus Duration (worker – elapsed time) – Disk IO, network transmission, parallel plan?

• Execution count – network roundtrip? • Plan cost – Parallelism – High volume of quick queries is bad, so is excessive DOP

• Index – current rows, rows at time stats generated, sample rows & date

Execution Plans estimate - actual • Actual: estimated cost, actual rows, DOP – Compile parameters – Actual rows/executions versus estimated

Execute stored procedure once for each possible code path – with appropriate parameters

Execution Plans Analysis • Predicate – index key columns does not matching full SARG – SQL has function on SARG, data type mismatch – Compile parameters & statistics – Actual and Estimated rows/execution mismatch – Large table scans: how many rows output? – Rebinds and Rewinds – key lookup – Parallelism

Execution Plans • Pay attention to: – Compile parameters – Large table scans: how many rows output? – Predicate • search condition without suitable index

– Rebinds and Rewinds – key lookup – Parallelism

Index Usage – missing IX, excess IX? • Index usage – seek, scan, lookup & update – Unused indexes (infrequent code?) can be dropped

– Infrequent usage: check plan references – Similar indexes (leading keys) • Same keys, different order • Check plan reference – consolidate if possible

• Scans to large tables or even nonclustered IX – Is it real (SELECT TOP 1 may not be a real scan)

• Lookups – can these be reduced?

SQL Server Skills & Roles Data Architect normalization

Architect Table structure, unique keys

Developers SQL code Performance DBA Index + Statistics Maintenance Hardware & Storage

• .

SQL Server Performance History • Before DMVs (SQL Server 2000) – Profiler/Trace to get top SQL – Execution plans – not really exportable – Which indexes are actually used?

• Today – Trace/Extended Events sometimes not necessary • If the dm_exec_query_stats content is good

– Execution plans are exportable – Index Usage Stats

How much can be automated? • Data collection

all, of course

– Top resource consumers, etc.

• Assessment

sometimes

– Is there a problem – Can it be fixed or improved

• Fix/Change – Indexes – SQL – sometimes – Table structure, architecture If problems could be solved by pushing a button, what would be the skill requirements to be a DBA?

sometimes

no Great accomplishments – 99% perspiration 1% inspiration

Performance Approaches • Check against list of “Best Practices” • Manual DMV scripts approach – Find Top 5 or 10 SQL – Fix it if/when there is a problem

• All Indexes and procedures/SQL – Examine the complete set of stored procedures – Or the full list of SQL statements – Good indexes for all SQL, no more indexes than

Why bother when there are no problems? • No problems for over 1 year – Never bothered to collect performance baseline

• Problem Today – Find it with DMV, fix it – the problem was xxx – but why did it occur today & not before? • Probably statistics or compile parameters, but prove it?

• Why ExecStats – SQL scripts? – too much manual work – Third party tools? – only find problem

Rigorous Optimization • Table structure, SQL, Client-side

• Cluster Key • Good (nonclustered) Indexes – All indexes are actually used

• No more indexes than necessary – Consolidate similar indexes • same keys, same order, or reverse order?

– What SQL is impacted? • Statistics update • Index maintenance

Must consider the full set of SQL/procedures in removing indexes?

SQL versus programming languages • SQL – great for data access – Not good for everything else – When SQL becomes horribly complicated – What would the code looks like in VB/Java/Cxx

Client-side program C#

Performance Information • Server, Storage • OS & SQL Server Settings • SQL Server – SQL, query execution statistics, execution plan – Compile parameters – Indexes and index usage statistics – Statistics sampling – when? percentage? skew?