Correctness invariants and architectural guidance for the adapter layer, coordinator, pgwire, peek paths, and timestamp oracle. Trigger when the user works on or asks questions about these subsystems — including "how does the coordinator work", "what are read holds", "explain the peek path", "how does timestamp selection work", "why does this query block". Also trigger when editing files in src/adapter/, src/pgwire/, or src/timestamp-oracle/.
When working on the adapter layer or related crates, always read and follow the
guidance in doc/developer/guide-adapter.md before making changes. That
document contains correctness invariants, architectural notes, and a list of
optimizations that have been tried and rejected for correctness reasons.
Read doc/developer/guide-adapter.md now.
When working on the timestamp oracle (src/timestamp-oracle/), also read the
module-level docs in src/timestamp-oracle/src/lib.rs and
src/timestamp-oracle/src/batching_oracle.rs for the TimestampOracle trait
contract and the batching/linearizability invariants.
When you or the user discover that an approach is incorrect or heading in a
wrong direction, suggest distilling the learning into
doc/developer/guide-adapter.md. Add it to whichever section fits best — a
new correctness invariant, a new subsection under an existing invariant, a
rejected optimization, or a new top-level section if needed. Describe what was
tried, why it's wrong, and any relevant context. This keeps the guide up to
date and prevents future re-discovery of the same pitfalls.
PyTorch深度学习模式与最佳实践，用于构建稳健、高效且可复现的训练流程、模型架构和数据加载。