test_scan_execution¶

The streaming scan runner against the in-memory FakeTransport: the ping-pong protocol (region alternation, exact INPUT_LENGTH, padded DMA), the overlap guarantee (a job only completes once the next batch’s DMA landed), and the error paths.

Attributes¶

`REGION_A`
`REGION_B`

Classes¶

`OverlapTransport`	JOB_STATUS reports DONE only once the next batch's H2C write has
`ResidentTransport`	DONE as soon as a job starts; records what was written.

Functions¶

`_batches`(rows_per_batch[, row_bytes]) → list[dau_driver.scan_execution.ScanBatch]
`test_run_scan_batches_alternates_regions_and_overlaps_staging`(...)
`test_run_scan_batches_accepts_a_lazy_iterator`(tmp_path)
`test_run_scan_batches_rejects_payload_row_mismatch`(...)
`test_run_scan_batches_propagates_job_errors`(tmp_path)
`test_run_scan_batches_validates_configuration_and_empty`(...)
`test_run_scan_batches_surfaces_staging_failures_promptly`(...)
`test_run_scan_batches_rejects_zero_row_batches`(tmp_path)
`_record_batches`(records, *, batch_count)
`_decoded_records`(batch)
`test_key_partitioned_batches_confine_skewed_keys_and_balance_within_largest_group`()
`test_key_partitioned_batches_degenerate_cases`()
`test_key_partitioned_batches_delay_input_consumption_and_encode_payloads_one_at_a_time`()
`test_key_partitioning_avoids_positional_split_state_hazard`()
`test_run_scan_batches_refuses_positional_batches_for_per_key_state`(...)
`test_run_scan_batches_accepts_key_partitioned_batches_for_per_key_state`(...)
`test_batch_balance_is_bounded_by_the_largest_key`() → None	A key cannot be split without resetting the per-key state the
`_encode_pairs`(records) → bytes
`test_key_range_batches_bound_the_distinct_keys_per_batch`() → None
`test_key_range_batches_put_each_key_in_exactly_one_batch`() → None	The property that makes combining per-batch group records a
`test_key_range_batches_preserve_arrival_order_within_a_batch`() → None
`test_key_range_batches_handle_the_boundaries`() → None
`test_key_range_batches_merge_back_to_the_one_shot_answer`() → None	End to end against the tile golden: capacity-bounded batches run
`test_a_resident_run_uploads_nothing`(tmp_path) → None	The whole point: rows already on the device are queried without a DMA.
`test_a_resident_run_still_programs_the_exact_input_length`(...) → None	Address and length are the contract; getting either wrong reads the
`test_a_misaligned_resident_address_is_refused`(tmp_path) → None	An unaligned DMA region is a silent-wrong-bytes failure on silicon,

Module Contents¶

test_scan_execution.REGION_A = 0¶

test_scan_execution.REGION_B = 134217728¶

class test_scan_execution.OverlapTransport(total_batches: int)¶

Bases: dau_driver.testing.FakeTransport

JOB_STATUS reports DONE only once the next batch’s H2C write has landed (the last job completes unconditionally): a serial stage-after-finish runner deadlocks into the timeout here, so a green run proves staging overlaps the running job.

total¶

started = 0¶

write_u32(offset: int, value: int) → None¶

read_u32(offset: int) → int¶

test_scan_execution._batches(rows_per_batch: list[int], row_bytes: int = 16) → list[dau_driver.scan_execution.ScanBatch]¶

test_scan_execution.test_run_scan_batches_alternates_regions_and_overlaps_staging(tmp_path)¶

test_scan_execution.test_run_scan_batches_accepts_a_lazy_iterator(tmp_path)¶

test_scan_execution.test_run_scan_batches_rejects_payload_row_mismatch(tmp_path)¶

test_scan_execution.test_run_scan_batches_propagates_job_errors(tmp_path)¶

test_scan_execution.test_run_scan_batches_validates_configuration_and_empty(tmp_path)¶

test_scan_execution.test_run_scan_batches_surfaces_staging_failures_promptly(tmp_path)¶

test_scan_execution.test_run_scan_batches_rejects_zero_row_batches(tmp_path)¶

test_scan_execution._record_batches(records, *, batch_count)¶

test_scan_execution._decoded_records(batch)¶

test_scan_execution.test_key_partitioned_batches_confine_skewed_keys_and_balance_within_largest_group()¶

test_scan_execution.test_key_partitioned_batches_degenerate_cases()¶

test_scan_execution.test_key_partitioned_batches_delay_input_consumption_and_encode_payloads_one_at_a_time()¶

test_scan_execution.test_key_partitioning_avoids_positional_split_state_hazard()¶

test_scan_execution.test_run_scan_batches_refuses_positional_batches_for_per_key_state(tmp_path)¶

test_scan_execution.test_run_scan_batches_accepts_key_partitioned_batches_for_per_key_state(tmp_path)¶

test_scan_execution.test_batch_balance_is_bounded_by_the_largest_key() → None¶: A key cannot be split without resetting the per-key state the partitioning exists to protect, so one dominant key caps how evenly batches can balance — and therefore how much overlap can hide. Pinned because the balancing heuristic could otherwise be read as a promise it cannot keep under skew.

test_scan_execution._encode_pairs(records) → bytes¶

test_scan_execution.test_key_range_batches_bound_the_distinct_keys_per_batch() → None¶

test_scan_execution.test_key_range_batches_put_each_key_in_exactly_one_batch() → None¶: The property that makes combining per-batch group records a concatenation rather than a summation.

test_scan_execution.test_key_range_batches_preserve_arrival_order_within_a_batch() → None¶

test_scan_execution.test_key_range_batches_handle_the_boundaries() → None¶

test_scan_execution.test_key_range_batches_merge_back_to_the_one_shot_answer() → None¶: End to end against the tile golden: capacity-bounded batches run through the hash-mode model and merge to what one unbounded pass would have produced.

class test_scan_execution.ResidentTransport¶

Bases: dau_driver.testing.FakeTransport

DONE as soon as a job starts; records what was written.

starts = 0¶

write_u32(offset: int, value: int) → None¶

read_u32(offset: int) → int¶

test_scan_execution.test_a_resident_run_uploads_nothing(tmp_path) → None¶: The whole point: rows already on the device are queried without a DMA. On the measured platform the feed is the larger term of the device path (40.0 ms total against a ~10.4 ms job), so removing it from the loop is what amortization means.

test_scan_execution.test_a_resident_run_still_programs_the_exact_input_length(tmp_path) → None¶: Address and length are the contract; getting either wrong reads the wrong bytes and would be indistinguishable from a bad answer.

test_scan_execution.test_a_misaligned_resident_address_is_refused(tmp_path) → None¶: An unaligned DMA region is a silent-wrong-bytes failure on silicon, so it is refused where the mistake is still cheap.