A Critique of Snapshot Isolation

"A Critique of Snapshot Isolation" is a notable transaction paper implemented in CockroachDB and Badger. It analyzes snapshot isolation (SI) and highlights that while SI offers high concurrency by checking only for write-write conflicts, it fails to guarantee serializability, allowing anomalies like write skew. The authors introduce write-snapshot isolation (WSI), an alternative that ensures serializability by preventing read-write conflicts rather than write-write conflicts. The paper demonstrates that WSI can achieve performance comparable to SI while providing the stricter consistency guarantees of serializability, making it a more robust solution for transactional support in large-scale distributed databases.

Transcript

1. A critique of snapshot isolation Authors: Maysam Yabandeh , Daniel Gómez Ferro - Yahoo Research Presenter: Arjun Sunil Kumar
2. 1. Isolation Levels and anomalies 2. Serialization Strategies 3. SI and SSI Implementations 4. Intuitive example 5. Eval Agenda
3. Isolation Levels
4. 1983 - ACID - Txn Property https://blog.bytebytego.com/p/ep89-what-does-acid-mean
5. Reference: https://learnsql.com/blog/history-of-sql-standards/ https://dev.mysql.com/doc/refman/8.0/en/compati bility.html 1986 : History of SQL Standards
6. SQL 92 - Transaction Isolation RedBook - http://www.redbook.io/ch6-isolation.html SQL 92 - https://www.contrib.andrew.cmu.edu/~shadow/sql/sql1992.txt
7. Isolation Levels and Anomalies https://pubs.opengroup.org/onlinepubs/9695959099/toc.pdf
8. 1. Dirty Read Good for Write Heavy Workload - Metrics Ref: https://www.thenile.dev/blog/transaction-isolation-postgres Read Uncommitted No locks
9. 3. Non-Repeatable / Fuzzy Reads R W Ref: https://www.thenile.dev/blog/transaction-isolation-postgres Read Committed (W: Locked, R: Early release) W
10. 4. Phantom Reads Ref: https://www.thenile.dev/blog/transaction-isolation-postgres R R I Repeatable read (W: Lock, R: Lock, No Range Lock)
11. 6. No Anomaly Serializable (R/W/Range Lock) T1 T2 R(x) W(x) R(y) W(y) R(x) W(x) R(y) W(y) Range Locks applies for SELECT … WHERE SELECT … BETWEEN 2PL - Growing Phase: Get necessary Shared/Exclusive/Range locks on the dataset. - Do work - Shrinking Phase: release locks
12. 1995 - A Critique of ANSI SQL Isolation Levels https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-95-51.pdf
13. 1999 - Weak Consistency: A Generalized Theory and Optimistic Implementations for Distributed Transactions - Atul Adya
14. 2. Lost Update & Cursor Stability https://www.scaler.com/topics/lost-update-problem-in-dbms/ R R W W
15. Snapshot Isolation - a form of MVCC https://arxiv.org/pdf/2405.18393 https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-95-51.pdf “Preventing Lost Updates” → WW Conflict
16. Snapshot Isolation - a form of MVCC S S C C https://unsplash.com/photos/group-of-men-running-in-track-field-HgoKvtKpyHA
17. 5. Write Skew R same R same W diff W diff Ref: https://drkp.net/papers/ssi-vldb12.pdf
18. Fixing Snapshot Isolation - RW Conflict https://arxiv.org/pdf/2405.18393
19. Read Only Txn’s are freq.
20. Critique - WW vs RW Conflict
21. Read Snapshot Isolation vs WSI https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-95-51.pdf Read SI Write SI No Anomaly 1 2 3 4 5 6
22. Serialization Strategies (Don’t take it too seriously)
23. Isolations are tricky!
24. 1. Sequential Execution Main Memory Database Each Core runs Single Thread Exec Engine. Each Core has a single partition. https://www.vldb.org/pvldb/vol6/p1942-debrabant.pdf https://hstore.cs.brown.edu/documentation/architecture-overview/
25. 2. 2PC Ref: https://p-org.github.io/P/tutorial/twophasecommit/ Lock is part of record.
26. 1. Timeout 2. Pre-commit Phase 3. Higher Message Overhead 3. 3PC 1. Participant Failure - Halts Txn. 2. Coordinator Failure - Needs recovery time.
27. 4. 2PC with Paxos https://www.mydistributed.systems/2020/08/google-spanner.html Lock Table in Coordinator.
28. - Deadlock/Starvation - Failed Locks or Slow Txn prevents the others Txns from making progress. Problems With Locks Based Approach
29. Serializable - It’s like Git Commit https://fangpenlin.com/posts/2013/09/30/keep-a-readable-git-history/ SERIALIZABLE isolation guarantees that even though transactions may execute in parallel, the result is the same as if they had executed one at a time, without any concurrency. This ensures data correctness by preventing all "anomalies" allowed by weaker isolation levels.
30. Lock based Strategy vs Timestamp Based Strategy Grow Shrink Work LOCK (Eager) TS Validate Work OCC (Lazy) Commit Conflict Detection! May or May not use locks TS Work/Validate TO (As you go) Commit Validate Work Scheduling (Eager) Order TS Batch You have more chance of conflict.
31. 5. Basic Timestamp Ordering > Record keeps track of WTS, RTS https://www.britannica.co m/topic/Big-Ben-clock-Lo ndon https://15445.courses.cs.cmu.edu/fall2023/notes/17-timestampordering.pdf BTO uses timestamps to control the access and update of data items at the time of execution. It aborts any transaction that violates the timestamp order. OCC uses timestamps to validate the transactions at the end of execution. Ref: https://muratbuffalo.blogspot.com/2022/11/timestamp-based-algorithms-for.html
32. 6. OCC Conflict Detection - Precedence Graph https://drkp.net/papers/ssi-vldb12.pdf https://slideplayer.com/slide/6247922/
33. 6. OCC - Precedence Graph Cont. https://drkp.net/papers/ssi-vldb12.pdf
34. 7. OCC - Dangerous Structure https://drkp.net/papers/ssi-vldb12.pdf inConflict:0 outConflict:1 inConflict:1 outConflict:0 inConflict:0 outConflict:0 T2 T1 T3
35. 7. OCC - Dangerous Structure - Cont. https://courses.cs.washington.edu/courses/cse444/08au/544M/READING-LIST/fekete-sigmod2008.pdf
36. 8. OCC - Central Oracle (Recent Committed Txns) Central Oracle for Ts and Last Txns. https://www.etsy.com/listing/60 3500543/cairn-photo-clock-sta cked-stones https://15445.courses.cs.cmu.edu/fall2023/notes/17-timestampordering.pdf
37. 9. Scheduling - Calvin Sequencer https://www.cs.yale.edu/homes/thomson/publications/calvin-sigmod12.pdf Smaller Timeunit batches Need to know R/W set before hand Preloading. https://blog.acolyer.org/2019/03/29/calvin-fast-distributed-transactions-f or-partitioned-database-systems/
38. 9. Scheduling - Calvin Sequencer Cont. https://blog.acolyer.org/2019/03/29/calvin-fast-distributed-transactions-f or-partitioned-database-systems/
39. Serialization Techniques - Summary Locks TSO Conflict Check at Work Graph Testing For Conflict Check At End Central Manager For Conflict Check At End Batching and Scheduling Seq Exec
40. Snapshot Isolation Impls
41. Regular SI - Needs Start/Commit TS - Start Ts - Commit Ts
42. Overlaps Txi Txj Row Si Ci Sj Cj Spatial Temporal
43. 1. Lock Based SI Stores latest Commit TS for WW conflict check. Used by lock-free SI. Used in 2PC
44. 2. Lock Free SI https://www.etsy.com/listing/60 3500543/cairn-photo-clock-sta cked-stones
45. Algorithm Si Ci Sj Cj Hit’s DB
46. 3. Lock Free Serializable SI Txn doesn’t commit if read set is modified.
47. Conditions Txi Txj Si Ci Sj Cj Row W : {v1} R: {v1, v3…}
48. Algorithm RSI WSI
49. Traffic Types 1. RO Read Set = empty Write Set = empty Analytics 2. OLTP Read Set = small ~20 rows Write Set = small ~20 rows • Approximate Data Set • Separate OLAP workload from OLTP oracle.
50. Impl in Big Data Landscape
51. Only stores last NR committed rows https://www.etsy.com/listing/60 3500543/cairn-photo-clock-sta cked-stones
52. Bookkeeper WAL
53. Oracle https://github.com/SarthakMakhija/serialized-snapshot-isolation https://github.com/dborchard/tiny-txn https://github.com/dgraph-io/badger
54. Txn
55. Read Ts 1. Get the next timestamp 2. Add to minHeap so that we can find the smallest active readTs for GC. 3. Wait for all ready to commit txn to commit.
56. Read Op
57. Write Op
58. Commit Op
59. Commit Ts 1. Conflict Check 2. Close ReadTs 3. GC Commit Txn’s older than lowest ReadTs in MinHeap. 4. Generate CommitTs 5. Add commitTs to Commit Visibility Waiter
60. Conflict Check
61. GC Find the minTs last read
62. Understanding Core Idea
63. The Cashier A →$10→ B A →$20→ C 3 4 5 6 Oracle Compute Nodes Performing Txns Txn Req
64. Reat Ts 3 4 5 6 A →Balance 7 Txn’s which have CommitTS (CommitWaiter)
65. Commit Wait A →Balance 7
66. Snapshot A →Balance 7 Guarantees snapshot you take, has seen all the Txns before the readTs
67. Do Work 8 9 A →$10→ B From/To 7
68. Committed Txns Pile Up A →$10→ B From/To 7 8, 9
69. Start Commit A →$10→ B 7 8, 9 S:7
70. Conflict Detect A →$10→ B 7 8, 9 S:7
71. Commit Ts A →$10→ B 7 8, 9 C:10
72. Read Done/GC Trigger A →$10→ B 7 8, 9 C:10 GC cheques before timestamp 7, If 7 is the MinHeap Peek. Kind of like Sliding Window of Committed Txn’s
73. Executor Apply / Commit Close A →$10→ B 8, 9,10
74. Commit Success A →$10→ B 8, 9,10
75. Eval
76. Expectation • RO Txn’s are never aborted by either RSI or WSI • RSI and WSI have same level of concurrency
77. Workload
78. Evaluation Aims to Answer
79. ReatTs(0.17ms) and CommitTs(4.1ms) Latency
80. Overhead on Status Oracle - W vs RW set CS is more for WSI 2x Memory Item Write Set Read Set Clients: 1 to 64 (exponential) To get higher throughput, you could partition the database and use a status oracle for each partition.
81. Overhead/Perf with HBase Client # 5, 10, 20, 40, 80, 160, 320, 640
82. Abort Rate Abort Rate in ZipfianLatest increases more quickly. ZipfianLatest, the read set is selected mostly from the recent written data, which increases the chance of R-w conflict in WSI.
83. To Summarize • Serializability is Expensive in 2PL • Weaker Isolation Levels and it’s anomaly • Snapshot Isolation and Write Skew Anomaly • Use RW conflict instead of WW conflict for Serializability • Doesn’t impact RO Txn. • WSI and RSI performance almost comparable. • Serializability could be brought to lock-free txn systems without hurting Performance.
84. Thank you! Icons: - flaticon.com 84
85. RSI vs WSI RSI WSI Serializability Read from a particular snapshot readTs Write to a Particular Snapshot (ie Serializable in Timeline) Take the Dr Alice, Dr Bob situation.