Distributed ACID Transac2ons in Apache Ignite

Size: px

Start display at page:

Download "Distributed ACID Transac2ons in Apache Ignite"

Reginald Morgan Roberts
6 years ago
Views:

1 Distributed ACID Transac2ons in Apache Ignite Akmal Chaudhri GridGain hbp://ignite.apache.org #apacheignite

2 My Background Pre-2000 Developer Academic (City University) Consultant Technical Architect Post-2000 Senior Architect Senior IT Specialist Technical Instructor Solu<ons Architect Independent Consultant

3 My Background Broad industry experience Worked with various technologies Programming languages IDE Database systems Client-facing roles Developers Senior execu<ves Journalists Publica<ons and presenta<ons

4 4

5 Agenda Apache Ignite Transac<ons Overview Concurrency Modes and Isola<on Levels Two-Phase Commit (2PC) Protocol One-Phase Commit (1PC) Op<miza<on RAM and Disk Level Consistency Deadlock Handling Demos Q & A

6 Apache Ignite Transac2ons Overview ACID-compliant Distributed Integrated with JTA Used by Financial Ins<tu<ons Atomicity modes Atomic Transac<onal

7 Atomicity Modes Atomic mode Updates are applied right away No locks => beber performance Transac<onal mode Opera<ons meets transac<on boundaries Fully ACID-compliant

can be used simultaneously try (Transaction tx = transactions.

8 Concurrency Modes and Isola2on Levels Concurrency modes for transac<ons Pessimis<c Op<mis<c Isola<on levels READ_COMMITTED REPEATABLE_READ } SERIALIZABLE All combina<ons can be used simultaneously try (Transaction tx = transactions.txstart()) { Integer hello = cache.get("hello"); if (hello == 1) cache.put("hello", 11); cache.put("world", 22); tx.commit();

$Pessimis2c Transac2ons try (Transaction tx = Ignition.ignite().transactions().txStart(PESSIMISTIC, REPEATABLE_READ)) { Account acct = cache.$

9 Pessimis2c Transac2ons try (Transaction tx = Ignition.ignite().transactions().txStart(PESSIMISTIC, REPEATABLE_READ)) { Account acct = cache.get(acctid); assert acct!= null; // Deposit into account. acct.update(amount); // Store updated account in cache. cache.put(acctid, acct); } tx.commit();

get(acctid); assert acct!= null; // Deposit into account. acct.update(amount); // Store updated account in cache.

10 Op2mis2c Transac2ons while (true) { try (Transaction tx = ignite.transactions().txstart(transactionconcurrency.optimistic, TransactionIsolation.SERIALIZABLE)) { Account acct = cache.get(acctid); assert acct!= null; // Deposit into account. acct.update(amount); // Store updated account in cache. cache.put(acctid, acct); } tx.commit(); break; } catch (TransactionOptimisticException e) { // Transaction has failed. Retry. }

Performance Considera2ons: Put vs PutAll 1 4 6 Network

11 Performance Considera2ons: Put vs PutAll Network Operations Node A Sequential Put Node B Node C

12 Performance Considera2ons: Put vs PutAll Network Operations Node A Putall 2 5 Node B 3 6 Node C

13 Two-Phase Commit protocol Phase 1 - Prepare Coordinator Prepare Prepare Prepare Prepare

14 Two-Phase Commit protocol Phase 2 - Commit Coordinator Commit Commit Commit Commit

15 Two-Phase Commit: Benefits Advantages of in-memory In-memory cache is vola<le The only reply for the "Prepare" is "Yes" Expands to the disk level Ignite Persistent Store 3rd party Persistence

16 Two-Phase Commit: Backup Node Failures Prepare Coordinator Prepare Prepare Prepare

17 Two-Phase Commit: Primary Node Failures Prepare Coordinator Prepare Prepare Prepare

18 Two-Phase Commit: Primary Node Failures Prepare Coordinator Prepare Prepare

19 Two-Phase Commit: Primary Node Failures Commit or Rollback? Commit or Rollback? Commit Coordinator Commit Commit Commit

20 Two-Phase Commit: Coordinator Failure Recovery Protocol Commit Commit Coordinator Did you Commit? Did you Commit? Did you Commit? Did you Commit?

21 One-Phase Commit Protocol 1PC is 2PC op<miza<on For best performance Minimize cluster nodes in a transac<on All transac<on entries belong to same par<<on Backup copies will also be grouped together Custom affinity mapping Example All employees working for the same company will be mapped to the same par<<on using companyid

22 One-Phase Commit (1PC) Coordinator Primary Node Backup Node Keys 1, 4 and 7 map to the same partition start transaction acknowledge lock key 1 lock key 4 lock key 7 commit backups for keys 1, 4, 7 acknowledge commit commit primary copies

23 One-Phase Commit: Primary Node Failure Coordinator Primary Node Backup Node Keys 1, 4 and 7 map to the same partition start transaction lock key 1 lock key 4 lock key 7

24 One-Phase Commit: Primary Node Failure Coordinator Primary Node Backup Node Keys 1, 4 and 7 map to the same partition start transaction lock key 1 lock key 4 lock key 7 commit backups for keys 1, 4, 7 acknowledge commit

25 One-Phase Commit: Primary Node Failure Recovery Protocol Coordinator Primary Node Backup Node Keys 1, 4 and 7 map to the same partition start transaction lock key 1 lock key 4 lock key 7 commit backups for keys 1, 4, 7 acknowledge commit commit primary copies

26 RAM and Disk Level Consistency Data are stored both in RAM and on disk ACID-compliant on all the levels 2PC Protocol is expanded to disk Ignite Persistent Store 3rd Party Stores

27 Durable Memory Architecture Memory split into regions Regions split into segments Segments split into pages Free-Lists track free space Pages can store: Data Indexes Metadata

datasets ACID-Compliant Write-ahead logs Stores superset of data

28 Ignite Persistent Store Expands Virtual Memory to Disk Flash, SSD, Intel 3D Xpoint ANSI-99 SQL-Compliant Join RAM and disk datasets ACID-Compliant Write-ahead logs Stores superset of data If a page is in RAM, it is always on disk Same format Instantaneous Restarts

Provides recovery mechanism Checkpoin<ng Triggered

29 Durability and Consistency Write-Ahead Log (WAL) Updates appended to node s WAL WAL propagates updates to disk Provides recovery mechanism Checkpoin<ng Triggered periodically based upon config Copy dirty pages from RAM to disk Reduces WAL size

30 3rd Party Stores: Consistency Coordinator writes to the database first Commits in the cluster a8erwards The database must be transac<onal RDBMS

31 3rd Party Stores: 2PC Protocol Phase 1 - Prepare Coordinator 2. Prepare 2. Prepare 2. Prepare 2. Prepare

32 3rd Party Stores: 2PC Protocol Phase 2 - Commit 3. Commit 3. Commit Coordinator 1. Commit 3. Commit Database 3. Commit

33 3rd Party Stores: Failures Commit Coordinator Commit Reload Reload Reload Database

34 Deadlock Handling Deadlock Detec<on Helps prevent distributed deadlocks Facilitates deadlock debugging and avoidance Deadlock-free Transac<ons Op<mis<c and Serializable mode Avoids possibility of deadlock One of the transac<ons always fails Failed transac<on can be retried

35 Demos

36 ANY QUESTIONS? Thank you for joining us. Follow the conversa<on. hbp://ignite.apache.org #apacheignite

37 Resources Two-Phase-Commit for Distributed In-Memory Caches hbp://gridgain.blogspot.co.uk/2014/09/two-phasecommit-for-distributed-in.html Two-Phase-Commit for In-Memory Caches - Part II hbp://gridgain.blogspot.co.uk/2014/09/two-phasecommit-for-in-memory-caches.html One-Phase-Commit - Fast Transac<ons For In-Memory Caches hbp://gridgain.blogspot.co.uk/2014/09/one-phasecommit-fast-transac<ons-for.html

GridGain and Apache Ignite In-Memory Performance with Durability of Disk

GridGain and Apache Ignite In-Memory Performance with Durability of Disk Dmitriy Setrakyan Apache Ignite PMC GridGain Founder & CPO http://ignite.apache.org #apacheignite Agenda What is GridGain and Ignite