2026-03-24 Test Coverage Gap Closure

• Status: implemented
• Date: 2026-03-24
• Baseline repo commit before this note's documented state: a30f01a93d0f9e2136033f2498618dab236a01a6
• Baseline commit summary: test: update async reader tests to use timedelta for delays
• This coverage note was recorded in commit:
• c1fc90f714c7cd13d0daf9a077f1394a5adaf69c docs: reflect timedelta and renamed parameters in documentation

Summary

This note records a focused test-coverage closure pass that added 73 tests across 12 files, targeting states and error paths that had previously been uncovered.

The emphasis was not broad percentage improvement. It was concentrated coverage in areas where missing tests could hide correctness, lifecycle, or error-propagation bugs.

What Was Added

73 new tests across 12 files, organized by severity.

Critical — Data Corruption / Contract Violations

test_publish_error_paths.py (7 tests)

shardyfusion/_writer_core.py:publish_to_store()

The two-phase publish (manifest PUT → _CURRENT pointer PUT) had its entire error-handling path marked # pragma: no cover. These tests cover:

• PublishCurrentError raised by store.publish() with a manifest_ref — triggers the retry loop that calls store.set_current() directly
• Retry succeeds on the 2nd attempt (first set_current fails, second succeeds)
• All 3 set_current retries exhausted — raises PublishManifestError
• PublishCurrentError with manifest_ref=None — re-raised immediately (no recovery possible)
• Generic exception during store.publish() — wrapped in PublishManifestError
• Happy-path: MANIFEST_PUBLISHED metric emitted on success

test_cel_routing_reader.py (6 tests)

shardyfusion/reader/reader.py and concurrent_reader.py

All three reader types accept routing_context for CEL sharding, but no reader test had ever used CEL. These tests cover:

• ShardedReader.get() and route_key() with a CEL key-only manifest (key % 4)
• ConcurrentShardedReader.route_key() with CEL
• ShardedReader.get() / multi_get() / route_key() with multi-column CEL + boundaries (region → bisect_right shard selection)
• Calling route_key() without routing_context on a multi-column CEL manifest — raises ValueError because the router cannot function without the extra columns

test_adapter_lifecycle.py (5 tests)

shardyfusion/_shard_writer.py:write_shard_core()

No test had simulated an adapter that fails mid-write. These tests cover:

• adapter.flush() raises OSError — write_shard_core wraps it in ShardWriteError
• adapter.checkpoint() returns None — shard result is written but checkpoint_id is None
• adapter.__exit__() raises — error propagates out
• Factory __call__ raises ConnectionError — propagates, possibly wrapped
• Zero rows written — row_count == 0, no crash

High — Reliability / Correctness

test_router_transitions.py (4 tests)

shardyfusion/routing.py, shardyfusion/reader/reader.py

When a reader refreshes, num_dbs or the shard layout can change. Tests cover:

• num_dbs increases (2→4) on refresh — snapshot_info().num_dbs reflects the new count
• num_dbs decreases (4→2) on refresh — route_key() uses the new modulus
• Manifest with missing db_ids, such as only shards 0 and 2 out of 3 — SnapshotRouter pads the gap with a synthetic null shard; shard_details() shows row_count=0 for shard 1
• Refresh from sparse to full manifest — previously-null shard now has real data

test_multi_get_failures.py (5 tests)

shardyfusion/reader/reader.py and concurrent_reader.py:multi_get()

When multi_get fans out to multiple shards and one fails, the error contract was unspecified. Tests cover:

• ShardedReader: one shard fails — exception raised, not silent partial result
• ConcurrentShardedReader: one shard fails — exception raised
• All shards fail — SlateDbApiError raised
• All keys on the same failing shard — original exception preserved
• ConcurrentShardedReader with max_workers=2 (executor path) — SlateDbApiError propagated from the failed future

test_async_deferred_cleanup.py (4 tests)

shardyfusion/reader/async_reader.py:_release_state()

When refresh() retires old state, cleanup is deferred via loop.create_task() until borrows drop to zero. Tests cover:

• Borrow held during refresh, then released — deferred cleanup task is scheduled
• close() called while a borrow handle is still outstanding — releasing the handle afterwards does not crash
• Two rapid refreshes — reader lands on the correct final manifest
• get() after refresh — returns data from the new state

test_value_spec.py (14 tests)

shardyfusion/serde.py:ValueSpec

ValueSpec was only tested indirectly via integration tests. These are dedicated unit tests:

• binary_col: None value → b""; bytearray → converted to bytes; str → UTF-8; unsupported type → ConfigValidationError
• json_cols: selected columns only; all columns; None value in column; nested dict; Unicode round-trips; missing column → None in output; description field
• callable_encoder: basic callable; encoder that raises; description uses __name__

Medium — Robustness

test_ordering.py (10 tests)

shardyfusion/ordering.py:compare_ordered()

Previously only exercised indirectly through boundary validation. Direct tests cover:

• Integers and strings: less-than, equal, greater-than
• Incompatible types ("a" vs 1) — raises ValueError with the caller's message
• None vs int — raises ValueError
• bytes and float comparisons

test_prometheus_completeness.py (15 tests) and test_otel_completeness.py (6 tests)

shardyfusion/metrics/prometheus.py, shardyfusion/metrics/otel.py

Several MetricEvent values handled by the collectors had no tests. Added coverage for:

• Handled events: SHARD_WRITE_COMPLETED (counter + histogram), BATCH_WRITTEN, READER_MULTI_GET (counter + histogram), S3_RETRY_EXHAUSTED, RATE_LIMITER_DENIED
• Silently ignored events, by design and not mapped to instruments: SHARDING_COMPLETED, MANIFEST_PUBLISHED, CURRENT_PUBLISHED, READER_INITIALIZED, READER_REFRESHED, READER_CLOSED, SHARD_WRITE_RETRIED, SHARD_WRITE_RETRY_EXHAUSTED — verified they do not raise

test_batch_errors.py (10 tests)

shardyfusion/cli/batch.py

Error-path coverage for the batch script runner:

• commands key is a string or int, not a list — ValueError
• Empty commands: [] — parses successfully, runs nothing
• Command dict missing op field — reported as an error
• get without key, multiget with empty or non-list keys, route without key — each reported as an error
• Reader raises during get — caught, reported as error, not a crash
• on_error: continue — all remaining commands still run after an error
• on_error: stop — execution halts after the first error; exactly one output line emitted

test_python_writer_edge_cases.py (3 tests)

shardyfusion/writer/python/writer.py:_write_parallel()

Edge cases for parallel mode not covered by the existing crash test:

• Empty record list — all shard results have row_count == 0
• Single shard (num_dbs=1) — one result with all rows
• 4 shards, 100 records — total row count across all shards equals 100