Transaction Concurrency-Control Foundations
Use this skill when you need shared vocabulary + decision lenses for transaction concurrency control before analyzing a specific protocol or trace.
This skill is most useful when you catch yourself asking:
- “What exactly is the anomaly / conflict here?”
- “What does the protocol need to prevent, and what metadata would prove it?”
- “Is this abort required for correctness or just conservative?”
Quick Reference (pick the next file to read)
| If you need… | Read | |---|---| | Core terms and what “commit/abort” means in histories | transaction-concepts.md | | A crisp definition of the bug pattern (lost update, dirty read, write skew, phantoms) | concurrency-control-problems.md | | A guarantee taxonomy (serializable vs SI vs strict serializable; opacity if relevant) | anomaly-and-isolation-taxonomy.md | | Dependency/conflict edge types (rw/wr/ww; anti-deps; visibility) | conflict-taxonomy.md | | Where a mechanism fits (locking vs TO vs OCC vs MVCC vs SSI vs hybrid) | protocol-family-overview.md | | Concrete protocol mini-specs (2PL/TO/OCC/MVCC/SSI/TicToc) | representative-protocols.md |
Router: from “protocol cues” to what to read
When you have a protocol description or trace fields but don’t know what family it is, route by cues.
A. Is the protocol primarily blocking or aborting?
- Blocking on conflicts, lock modes (S/X), deadlocks mentioned → go to 2PL in representative-protocols.md
- Nonblocking execution + commit-time validation → go to OCC (and often TicToc/SSI live here too)
B. Does it assign a fixed transaction timestamp early?
- Per-txn fixed TS(T), and objects store RTS/WTS → go to Basic Timestamp Ordering
- Per-txn snapshot/start_ts + multiple versions → go to MVCC / Snapshot Isolation
C. Do reads choose a version from a version chain?
- Yes (visibility rule, begin/end timestamps, “as of snapshot”) → go to MVCC
- No, single current version plus per-key metadata → go to TO / OCC / TicToc-like
D. Do you see read timestamp extension or “valid-through” metadata?
- Fields like rts (read-valid-through), extension checks, or “extend rts if unchanged” → go to TicToc
E. Do you see dependency / anti-dependency tracking?
- Explicit rw anti-deps, “dangerous structure”, serialization-graph edges → go to SSI
F. What’s the proof object?
- Locks prove exclusion → 2PL
- Per-key RTS/WTS prove timestamp order → TO
- Read-set version check proves stability → OCC
- Visibility from snapshot + version chain proves consistent reads → MVCC/SI
- Anti-dep edges / dangerous structure prove no cycles → SSI
- Per-key wts/rts interval + extend-if-unchanged proves read validity at chosen commit_ts → TicToc
Minimal Working Vocabulary (keep this tiny)
When analyzing a trace or protocol, keep a running record with these nouns and edges:
transactions: T1, T2, ...
objects/keys: x, y, ...
versions: x@wts=..., x@ver=...
operations: R(x@...), W(x@...), commit(T), abort(T)
dependencies:
T_i ->wr T_j (T_j read from T_i's write)
T_i ->rw T_j (T_i read a version that T_j later overwrote; anti-dep)
T_i ->ww T_j (T_j overwrote T_i on same object)
If you can’t name the version read/written, you’re still in “story mode,” not “proof mode.”
Checklist: classify a protocol decision
Use this checklist before labeling an abort “unnecessary” or a commit “unsafe”:
- Claimed guarantee: serializable? strict? SI? opacity?
- What the transaction read: object + which version/timestamp evidence.
- What changed by validation time: overwrite? timestamp extension? dependency edge?
- What order would make it safe: a serial order or a timestamp assignment.
- What evidence is missing: did the protocol lack history/versions to prove safety?
Common Pitfalls
- Mixing scopes: per-key counters and timestamps often cannot order different keys.
- Equating “row order” with “happened-before”: log order can be buffered/merged.
- Assuming abort == anomaly: aborts can be conservative proof failures, not unsafe histories.