How In-Memory Affects Database Design

Download Report

Transcript How In-Memory Affects Database Design

How In-Memory Affects Database Design
Louis Davidson
Certified Nerd
Who am I?
Been in IT for over 19 years
Microsoft MVP For 10 Years
Corporate Data Architect
Written five books on
database design
• Ok, so they were all versions
of the same book. They at least
had slightly different titles each time
• The book doesn’t yet cover this topic
Attention: There Is Homework
(lots of it)
• I can’t teach you everything
about In-Memory in 1 hour
• The code will be available, but
it is still very rudimentary
• It will get you started, but is
only just the tip of the iceberg
• Do lots of thinkin’ and testin’
before divin’ in
Presentation Philosophy
Introduction: What exactly is
In-Memory OLTP in SQL Server 2014?
• A totally new, revamped engine for data storage, co-located in
the same database with the existing engine
– Obviously Enterprise Only…
• Purpose built for certain scenarios
• Terminology can be confusing
–Existing tables: Home - On-Disk, but ideally cached In-Memory
–In-Memory tables: Home - In-Memory: but backed up by On-Disk
• If you have enough RAM, On-Disk tables are also in memory
• In-Memory is both very easy, and very difficult to use
–But the implementation is very very different
Design Basics
(And no, I am not stalling for time due to lack of material)
• Designing and Coding is Like the
Chicken and the Egg
I was first
As if…
–Design is what you do before coding
–Coding patterns can greatly affect design
–Engine implementation can greatly affect
design and coding patterns
• We will discuss how In-Memory
technologies affect the entire
development lifecycle
Design Basics - Separate your design mind into
three phases
1. Logical (Overall data requirements in a data model format)
2. Physical Implementation Choice (Indexes, Physical
Structures, etc)
3. Physical (Relational Code)
• Before the engine choice I always suggested 3 before 2
• We will look at each of these phases and how in-mem may
affect your design
Logical Design
(Though Not Everyone’s Is)
• This is the easiest part of the presentation
• You still need to model
–Entities and Attributes
–Uniqueness Conditions
–General Predicates
• As I see it, nothing changes…
Logical Data Model
Physical Implementation
(Or DBA stuff that I only slightly care about)
• Everything is different, and I am not here to cover these details…
• In-Mem data structures coexist in the database alongside OnDisk ones
• Data is housed in RAM, and backed up in Delta Files and
Transaction Logs
–Delta files are stored as filestream storage
–The transaction log is the same one as you are used to
• Tables and Indexes are extremely coupled
• MVCC (Multi-Valued Concurrency Control) used for all isolation
Physical Implementation Overview
Client App
TDS Handler and Session Management
No improvements in
communication stack,
parameter passing, result
set generation
10-30x more efficient
Reduced log bandwidth &
contention. Log latency
Compiled SPs
and Schema
Hekaton Engine for
Memory_optimized Tables
& Indexes
Proc/Plan cache for ad-hoc
T-SQL and SPs
Interpreter for TSQL, query
plans, expressions
Access Methods
Existing SQL
Buffer Pool for Tables &
SQL Server.exe
Table Filegroup
Checkpoints are
background sequential IO
Transaction Log
Data Filegroup
Physical Design
(No, let’s not get physical)
• Your physical design will almost certainly need to be affected
• So much changes, even just changing the table structure
• In this section, we will discuss:
–Creating storage objects
• Table creation
• Index creation (which is technically part of the table creation)
• Altering a Table’s Structure
–Accessing (Modifying/Creating) data
• No Locks, No Latches, No Waiting
• Using Normal T-SQL (Interop)
• Using Compiled Code (Native)
Creating Storage Objects - Tables
• The syntax is the same as on-disk, with a few additional settings
• You have a durability choices
– In-Mem Table: Schema_Only or Schema_and_Data
– Database level for transactions: Delayed (also for on-disk tables)
• Aaron Bertrand has a great article on this here:
• You also have less to work with...
– Rowsize limited to 8060 bytes (Enforced at Create Time)
• Not all datatypes allowed (LOB types,CLR,sql_variant, datetimeoffset, rowversion)
– No foreign keys
– No check constraints
– Limited unique constraints (just one unique index per table)
• Note: There are memory optimized temporary tables too: See Kendra Little’s article here:
• Every durable (Schema_and_Data) table must have a primary key
Dealing with Un-Supported Datatypes…
• Say you have a table with 10 columns, but 1 is not allowed in a
In-Memory table
• First: Ask yourself if the table really fits the criteria we aren’t
done covering
• Second: If so, consider vertically partitioning
• CREATE TABLE In_Mem (KeyValue, Column1, Column2, Column3)
CREATE TABLE On_Disk (KeyValue, Column4)
• It is likely that uses of disallowed types wouldn’t be good for the
OLTP aspects of the table in any case.
Creating Storage Objects - Index creation
• Indexes are linked directly to the table
– 8 indexes max per table due to internals
– Only one unique index allowed
– Indexes are never persisted, but are rebuilt on restart
• Two Index Types
– Hash
• Ideal for single row lookups
• Fixed size, you choose the number of hash buckets (approx 1-2 * # of unique values
– Bw Tree
• Best for range searches
• Very similar to a BTree index as you (hopefully) know it, but optimized for MVCC and pointer connection to
• String index columns must be a binary collation (case AND access sensitive)
• For more in-depth coverage
– check Kalen Delaney's white paper ...
– Or for an even deeper (nerdier?) versions: “Hekaton: SQL Server’s Memory-Optimized OLTP
Engine” or The Bw-Tree: A
B-tree for New Hardware Platforms (
A Taste of the Physical Structures
• A table with two hash indexes
• From Kalen’s Whitepaper:
Creating Storage Objects - Altering a Table
• The is the second easiest slide
in the deck
• No alterations allowed - Strictly
Drop and Recreate
– You can rename a table, which
makes this at east easier
Accessing the Data - No Locks, No Latches, No
• On-Disk Structures use Latches and Locks to implement
• In-Mem Use Optimistic-MVCC
• You have 3 Isolation Levels:
–Evaluated before, or when the transaction is committed
–This makes data integrity checking "interesting"
• Essential difference, your code now must handle errors
Concurrency is the #1 difference you will deal with
• Scenario1: 2 Connections - Update Every Row In 1 Million Rows
• Any Isolation Level
• On-Disk
–Either: 1 connection blocks the other
–Or: Deadlock
• In-Mem
–One connection will fail, saying: “the row you are trying to update has
been updated since this transaction started” EVEN if it never commits.
Another slide on Concurrency
(Because if I had presented it concurrently with the other one, you wouldn’t have liked that)
• Scenario2: 1 Connection Updates All Rows, Another Reads All
Rows (In an explicit transaction)
• On-Disk
–Either: 1 connection blocks the other
–Or: Deadlock
• In-Mem
–Both Queries Execute Immediately
–In SNAPSHOT ISOLATION the reader will always succeed
• Commits transaction BEFORE updater commits: Success
• Commits transaction AFTER updater commits: Fails
Accessing the Data - Using Normal T-SQL (Interop)
• Using typical interpreted T-SQL
• Most T-SQL will work with no change (you may need to add
isolation level hints)
• A few Exceptions
–TRUNCATE TABLE - This one is really annoying :)
–MERGE (In-Mem table cannot be the target)
–Cross Database Transactions (other than tempdb)
–Locking Hints
Accessing the Data using Compiled Code (Native)
• Instead of being interpreted, the stored procedure is compiled to
machine code
• Limited synax (Like programming with both hands tied behind your
• Allowed syntax is listed in what is available, not what isn't
• Some really extremely annoying ones:
– SUBSTRING supported; LEFT, RIGHT, not so much
– No Subqueries
– OR, NOT, IN, not supported in WHERE clause
• So you may have to write some "interesting" code
The Difficulty of Data Integrity
• With on-disk structures, we used constraints for most issues
(Uniqueness, Foreign Key, Simple Predicates)
• With in-memory code, we have to implement in stored
–Uniqueness on > 1 column set suffers from timing (If N connections are
inserting the same data...MVCC will let them)
–Foreign Key can't reliably be done because:
• In Snapshot Isolation Level, the row may have been deleted while you check
• In Higher Levels, the transaction will fail if the row has been updated
–Check constraint style work can be done in stored procedures for the
most part.
Problem: How to Implement Uniqueness on > 1
• CREATE VIEW Customers.Customers$UniquenessEnforcement
SELECT customerId, emailAddress, customerNumber
Customers.Customers$UniquenessEnforcement (emailAddress)
• Msg 10794, Level 16, State 12, Line 8
The operation 'CREATE INDEX' is not supported with memory optimized tables.
Problem: How to Implement Uniqueness on > 1
Column Set: Multiple Tables?
• Wow, that seems messy… And what about duplicate customerId
values in the two subordinate tables?
Problem: How to Implement Uniqueness on > 1
Column Set: Simple code
• You can’t…exactly. But what if EVERY caller has to go through the
following block:
• DECLARE @CustomerId INT
SELECT @CustomerId = CustomerId
FROM Customers.Customer
WHERE EmailAddress = @EmailAddress
IF @customerId is null… Do your insert
• This will stop MOST duplication, but not all. Two inserters can check at
the same time, and with no blocks, app locks, or constraints even
available, you may get duplicates.
• Remember the term: Optimistic Concurrency Control
When Should You Make Tables In-Memory Microsoft's Advice
• From
Implementation Scenario
Benefits of In-Memory OLTP
High data insertion rate from multiple
concurrent connections.
Primarily append-only store.
Unable to keep up with the insert workload.
Eliminate contention.
Reduce logging.
Read performance and scale with periodic
batch inserts and updates.
High performance read operations,
especially when each server request has
multiple read operations to perform.
Unable to meet scale-up requirements.
Eliminate contention when new data arrives.
Lower latency data retrieval.
Minimize code execution time.
Intensive business logic processing in the
database server.
Insert, update, and delete workload.
Intensive computation inside stored
Read and write contention.
Eliminate contention.
Minimize code execution time for reduced
latency and improved throughput.
Low latency.
Require low latency business transactions
which typical database solutions cannot
Eliminate contention.
Minimize code execution time.
Low latency code execution.
Efficient data retrieval.
Session state management.
Frequent insert, update and point lookups.
High scale load from numerous stateless
web servers.
Eliminate contention.
Efficient data retrieval.
Optional IO reduction or removal, when
using non-durable tables
Implementation Scenario
When Should You Make Tables In-Memory
Louis's Advice
• More or less the same as Microsoft's really (duh!)
• Things to factor in
–High concurrency needs/Low chance of collisions
–Minimal uniqueness protection requirements
–Minimal data integrity concerns (minimal key update/deletes)
–Limited searching of data (binary comparisons only)
–Limited need for transaction isolation/Short transactions
• Basically, the “hot” tables in a strict OLTP workloads...
The Choices I made
Louis has improved his methods for estimating performace, but your mileage will still vary.
Louis’ tests are designed to reflect only one certain usage conditions and user behavior, but
several factors may affect your mileage significantly:
How & Where You Put Your Logs
Computer Condition & Maintenance
CPU Variations
Programmer Coding Variations
Hard Disk Break In
Therefore, Louis’ performance ratings are a minimally useful tool for comparing the
performance of different strategies but may not accurately predict the average performance
you will get.
I seriously suggest you test the heck out of the technologies yourself using my code, your
code, and anyone else’s code you can to make sure you are getting the best performance
Model Choices – Logical Model
Model Choices – Physical Model
Model Choices – Tables to Make In-Mem
The Grand Illusion (So you think your life is
complete confusion)
• Performance gains are not exactly what you may expect, even when they are
• In my examples (which you have seen), I discovered when loading 20000 rows
(10 connections of 2000 each)
– (Captured using Adam Machanic's tool)
A. On-Disk Tables with FK, Instead Of Trigger - 0.0641 seconds per row - Total Time – 3:55
B. On-Disk Tables withOUT FK, Instead Of Trigger - 0.0131 seconds per row - Total Time – 2:44
C. In-Mem Tables using Interop code - 0.0091 seconds per row - Total Time 2:31
D. In-Mem Tables with Native Code - 0.0035 second per row - Total Time – 1:23
E. In-Mem Tables, Native Code, SCHEMA_ONLY – 0.0003 seconds per row - Total Time – 1:00
• But should it be a lot better? Don't forget the overhead... (And SQLQueryStress
has extra for gathering stats)
• Code Review As We have time