Comprehensive Indexing via Automated Execution Plan Analysis Exec

Comprehensive Indexing via Automated Execution Plan Analysis (Exec. Stats) Joe Chang jchang 6 @ 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 Exec. Stats http: //www. qdpma. com/Exec. Stats/SQLExec. Stats. html Download: http: //www. qdpma. com/Exec. Stats. Zip. html Blog: http: //sqlblog. com/blogs/joe_chang/default. aspx

What is not in this session • 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 - Not enthusiastic, prolific, indiscriminate collector of rules

Why this topic on Indexing? • Ideal – Good indexes, but no more than necessary • Reality Indexes can be created or dropped for specific operations – Too many nonclustered indexes • How did this happen? – Poor choice of cluster key (my opinion) – Indexes added one at a time for specific query • Instead of modifying similar existing index – DBA not inclined to remove indexes • Afraid to touch the cluster key

Going from Current to Good Indexes • Determine the better choice of cluster key – Can eliminate many nonclustered index by itself • Not hard to determine, but down time required • Drop indexes not used (over long period) • Consolidate indexes Simple – with leading keys in the same order – Determine if indexes with same keys in different order can be consolidated • Need to find SQL that uses each index? Execution plan links SQL to index usage Does this involve work/risk? I don’t like work/risk!

Notes • 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

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 keys & const Row estimate propagation errors DOP Memory Parallel plans API Server Cursors: open, prepare, execute, close? SET NO COUNT Information messages Indexes Query Optimizer Execution Plan Storage Engine Hardware Statistics Tables and SQL combined implement business logic Compile Natural keys with unique indexes, not SQL Sampling & Re -compute parameters & variables Recompile temp table / table variable Index & Stats Maintenance 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

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 Indexes, key columns, Include list, filter, XML, Columns store etc. DBCC SHOW_STATISTICS dm_db_index_usage_stats dm_db_index_operational_stats dm_db_index_physical_stats dm_exec_query_profiles 2014 Real time query progress? 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 R 2 starting with SP 2. last_updated, rows_sampled, steps, unfiltered_rows, modification_counter

Execution Plan maps SQL to Indexes SQL Execution Plan Indexes dm_exec_query_stats sql_handle & plan_handle dm_exec_sql_text dm_exec_query_plan dm_exec_text_query_plan dm_db_index_usage_stats

Performance Oriented Approach • 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?

Systematic Approach • Get full list of: – stored procedures : schema + name – Scalar Functions (FN) schema + name – Inline & Table Valued Functions (IF, TF) • Need parameter list – Triggers (should be obsolete, my opinion) • Generate execution plan – match to indexes – Alternative, maintain a list of SQL

Real World Example July 2014 3. 6 B rows, 3. 4 TB, 1. 5 TB data 1. 8 TB indexes Key tables have 21, 8, 14, 4 and 14 nonclustered indexes

Index Reduction Dec 2014 4. 6 B rows (30%), 2. 3 TB, 1. 7 TB data 0. 5 TB indexes Key tables have 6, 4, 3, 2 and 4 nonclustered indexes (some nonclustered indexes compressed)

Compression – base tables Dec 2014 4. 6 B rows, 2. 3 TB, 1. 7 TB data 0. 5 TB indexes May 2015 5. 7 B rows, 0. 99 TB, 0. 62 TB data 0. 36 TB indexes

Table – 21 nonclustered Indexes Note: All indexes are full rows, not filtered

Table with 21 NC Indexes Jul 2014 5 frequently used NC indexes of 21

Dec – 6 Nonclustered Indexes Jul 2014 5 frequently used NC indexes of 21 Dec 2014 4 frequently used, 2 Filtered IX, note lead column Infrequently used indexes could probably be removed by re-working the query

May 2015 – 2 NC used Jul 2014 5 frequently used NC indexes of 21 Dec 2014 4 frequently used, 2 Filtered IX, note lead column May 2015 3 frequently used NC, 1 disabled pending removal

Note on Filtered Index Strategy Query is SELECT xx FROM NARSplit WHERE Is. Active = 1 AND Cession. Id IN (list) Then the index strategy used is CREATE INDEX IX ON table (Cession. Id, xx, Is. Active) WHERE Is. Active = 1

Balancing Select vs Update In the above table, the nonclustered index with several included columns nearly eliminates key lookups One column, Is. Active, in the include list was frequently updated Removing that column reduces need for update maintenance Since a key lookup is needed anyways, may as well remove all include columns

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 – 648 GB data 230 GB indexes w/o compression – 174 GB data 185 GB indexes w/compression • Reduction in I/O even with SSD storage far outweigh compression overhead! System memory: 256 GB (220) Storage: Violin (NAND Flash)

Index reduction HDD Storage Compression & statistics Violin GDC

Index Theory - Locality • Example Database – 10 B rows, 80 bytes per row, 800 GB, 100 M pages • (100 M x 8 KB/page = 800 GB, 100 rows per page) • Suppose 1% of rows are active, i. e. 100 M rows – There could be 1 active row in each page (100 M) • Possible if each table were clustered on a row guid – Possible for active rows to be in only 1 M pages • All rows in each of these page happen to be active • Build the cluster key to tend toward 2 nd option

Index Key SARG + Group/Order SELECT Transaction. Type, SUM(Amount) FROM Table WHERE Report. Date = ‘value’ Selective SARG AND other SARGS GROUP BY Transaction. Type Grouping Index should lead with the key SARG, then Group or Order, less selective SARGs can be in Include List

Index Key SARG + Order By SELECT Transaction. Type, SUM(Amount) FROM Table WHERE Report. Date = ‘value’ Selective SARG AND other SARGS GROUP BY Transaction. Type Grouping Index should lead with the key SARG, then Group or Order, less selective SARGs can be in Include List

Index Example w & w/o Partitioning Without partitioning Required indexes leading with: 1) Report. Date - for grouping 2) Row. Id - for single row access With partitioning Partition as CREATE UNIQUE CLUSTERED INDEX UCX ON Table (Row. Id, Report. Date) ON psdate (Report. Date) Partition key is not lead key of cluster index Search on Row. Id must check each partition Oracle Skip-Scan would be nice

Exec. Stats

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_id, type, Create date, modify date, Number of references (Num. Ref) (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 @P 1 RECOMPILE (estimated plan)

About Exec. Stats • 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)

Exec. Stats 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 Exec. Stats – 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?