Convert empty timestamp (Timestamp(0, 0)) to current timestamp on insert

[richardsimmonds]

Fixes documentdb#513

Problem

Native MongoDB replaces an empty timestamp value Timestamp(0, 0) in a top-level field of an inserted document with the current timestamp. DocumentDB stored the empty timestamp literally:

// MongoDB
db.coll.insert({_id: 1, k: Timestamp(0, 0)});
db.coll.findOne({_id: 1});  // { "_id" : 1, "k" : Timestamp(1770764455, 1) }

// DocumentDB (before this change)
db.coll.findOne({_id: 1});  // { _id: 1, k: Timestamp({ t: 0, i: 0 }) }

Change

Adds ReplaceEmptyTimestampsWithCurrentTime to the insert preprocessing path (PreprocessInsertionDoc in pg_documentdb/src/commands/insert.c), which rewrites any top-level Timestamp(0, 0) field to the current time. Matching MongoDB semantics:

• only top-level fields are converted (not nested documents or arrays)
• the _id field is exempt and stored as-is
• non-empty timestamps are unchanged

The common case (no empty timestamps) takes a scan-only fast path and returns the original document without rewriting. The increment portion is populated from the clock's nanosecond component, consistent with the existing $currentDate update operator (HandleUpdateDollarCurrentDate).

Tests

New regression test insert_empty_timestamp_tests (added to basic_schedule) covering:

• single and multiple top-level empty timestamps via insert_one
• the batch insert command path
• _id exemption, nested-document and array non-conversion
• non-empty timestamps stored as-is
• validation that converted values fall within the clock window around the insert (would fail on the old code, which stores t: 0)

Self-review

Checked: symbol usage verified against headers (PgbsonWriterAppendValue, bson_iter_key_string_view, StringViewEquals, IdFieldStringView); clock_gettime usage mirrors the compiling pattern in bson_update_operators.c; unique regress collection id (1985100); DCO sign-off present; diff is minimal (one C function + tests + schedule line). Known limitation: full extension build not run locally — CI is the authoritative compile/regress gate.

[patty-chow]

Review Summary

Verdict: LGTM — recommend merge (pending CI), with two non-blocking follow-ups noted below.

Correctness

• Placement of the rewrite in PreprocessInsertionDoc is right: it runs after RewriteDocumentValueAddObjectId and before ComputeShardKeyHashForDocument / PgbsonGetDocumentId, so the shard key hash is computed over the converted value — the stored document and its hash stay consistent even if a timestamp field is part of the shard key.
• All insert entry points funnel through this function: batch path (DoMultiInsertWithoutTransactionId, insert.c:621) and single/retryable path (ProcessInsertion, insert.c:1003). The worker path (CallInsertWorkerForInsertOne → command_insert_worker) receives the already-preprocessed document, so conversion happens exactly once on the coordinator, and retryable writes return the retry record before preprocessing — no double-conversion or retry skew.
• Detection scan + fast-path return of the original pgbson for the common case is the right shape; no extra allocation when no empty timestamps are present.
• Semantics match documented MongoDB behavior: top-level only, _id exempt, non-empty timestamps untouched.

C safety

• No leaks or null derefs spotted. The writer-built pgbson is allocated in the caller's memory context with the same lifetime as the document RewriteDocumentValueAddObjectId already produces there (including inside the batch path's subtransaction PG_TRY block). Stack iterators, no manual frees needed.
• Symbols check out: IdFieldStringView via commands_common.h (already included), bson_iter_key_string_view/StringViewEquals via bson_core.h, clock_gettime mirrors the compiling usage in bson_update_operators.c:603.

Tests

• Good coverage: insert_one, batch insert command, multi-field conversion, _id exemption, nested/array non-conversion, non-empty passthrough. The clock-window assertion would genuinely fail on old code (stored t=0 falls outside the window), so this is a real regression test, not a tautology.
• Test helpers mirror the established test_update_currentDate pattern in bson_update_document_tests.sql (clock window + ±1s flake buffer). Unique collection id 1985100; schedule line reuses an existing parallel group — consistent with repo conventions.

Non-blocking follow-ups

1. Upsert path gap — replace-style upserts (UpsertDocument → InsertOrReplaceDocument in update.c) don't go through PreprocessInsertionDoc, so an upserted document with Timestamp(0,0) is stored literally. Native MongoDB converts on the upsert-insert path too. Worth a follow-up issue rather than scope-creeping this PR.
2. Increment semantics — MongoDB populates i with an ordinal counter (distinct value per replaced timestamp within a second); this PR stores tv_nsec, matching the existing $currentDate implementation's documented TODO. Consistent-with-codebase is the right call; just noting the known divergence.

DCO / scope

• Signed-off-by present on the single commit. Diff is minimal — one static function, call site, test, schedule line. No unrelated changes.

CI

• Status at review time: gateway tests in progress; the build/functional jobs (functional-pr-gate, image build+test, build-documentdb-local) show skipped on this fork — so the self-review's "CI is the compile gate" assumption may not hold here. Recommend confirming at least one compile-exercising job actually ran green before merge.

[patty-chow]

Placement is correct — conversion before ComputeShardKeyHashForDocument means the hash is computed over the converted value, keeping the stored document and shard key hash consistent.

One gap to track separately: replace-style upserts (UpsertDocument -> InsertOrReplaceDocument in update.c) don't funnel through PreprocessInsertionDoc, so an upserted document containing Timestamp(0,0) will still be stored literally. Native MongoDB converts on that path too. Suggest a follow-up issue rather than expanding this PR.

[patty-chow]

Non-blocking: native MongoDB sets i to an ordinal counter (so multiple empty timestamps replaced in the same second get distinct, ordered values); storing tv_nsec matches the existing $currentDate implementation (bson_update_operators.c:615, which carries the same TODO). Fine to keep consistent with the codebase — just flagging the known divergence so it's a deliberate choice.

[patty-chow]

Nice — this assertion is a genuine regression guard: old code stores t=0, which falls outside the clock window, so the test fails pre-fix. Helper pattern matches the established test_update_currentDate approach in bson_update_document_tests.sql.