Upstream SQL inventory¶

Open the searchable inventory. Download the per-line CSV files below.

The primary label answers whether the actual query passes StreamFusion's all-or-nothing gate. Every interior operator must run natively; rowwise source/sink boundaries and their transposes are allowed. One unsupported interior operator sends the entire query to Flink. Labels require passing upstream assertions and describe admission, not speedup or per-query runtime work. All in-scope streaming SQL gaps are targets.

Flink	Passing cases	Accelerated	Should be accelerated	Not accelerated
1.18	6,014	1,430	1,861	2,723
2.2	9,155	3,280	2,864	3,011

Each row remains one test invocation. Expand Query plans in the report or download the Flink 1.18 query-plan CSV and Flink 2.2 query-plan CSV for individual root verdicts. A test with both admitted and fallback queries remains a coverage target. Source/constant-only roots have no interior SQL computation. Observed roots can repeat during planning and are not counts of distinct SQL strings or completed executions; the existing observations do not associate SQL strings or individual translation failures with roots. Expected translation-only failures receive no admission credit.

Suites¶

Use the suite filter to inspect runtime and connector corpora separately. The primary label always concerns whole-query admission. connector_label, connector_category and connector_note retain the separate connector I/O targets; sql_label is an alias of the primary label. A host source or sink at an allowed perimeter does not make a fully admitted SQL query partial. Direct Java format tests and DataStream/legacy DataSet programs bypass SQL admission.

Suite	Flink	Passed	Accelerated	Should be accelerated	Not accelerated	Skipped
delta	2.2	4	2	1	1	0
formats	1.18	158	19	2	137	10
formats	2.2	175	19	2	154	10
kafka	1.18	72	51	21	0	0
kafka	2.2	86	65	21	0	0
orc	1.18	44	3	0	41	0
orc	2.2	46	3	0	43	0
paimon	1.18	71	12	1	58	4
paimon	2.2	265	95	18	152	0
parquet	1.18	8	3	3	2	0
parquet	2.2	8	6	0	2	0
runtime	1.18	5,661	1,342	1,834	2,485	25
runtime	2.2	8,571	3,090	2,822	2,659	48

The checked-in observations predate transparent stock RocksDB selection on 1.18 and the explicit 2.2 changelog-state admission guard. Their labels remain evidence of those recorded runs. In particular, the 805 recorded 1.18 backend blockers include 234 stock-RocksDB cases and 571 changelog-state cases; these are invocation counts, not distinct SQL features. Re-run the unchanged suites to measure the new coverage; do not relabel historical observations as newly accelerated.

Streaming coverage targets¶

Counts are passing test invocations, including parameter variants; the same missing feature can affect many cases. Categories overlap when a case has several blockers.

Category	Flink 1.18	Flink 2.2
scalar-expression	451	1,704
aggregation	336	514
state-backend	805	0
window	341	443
join	100	91
table-function	60	75
async-and-model-functions	0	71
pattern-matching	24	24
watermark	24	24
connector-boundary	19	27
rank-and-deduplication	17	23
data-type	12	19
ordering	8	10
changelog	0	9

Outside acceleration coverage¶

Reason	Flink 1.18	Flink 2.2
batch	1,780	1,968
validation-or-host-error	378	470
api-validation	203	252
source-or-constant-only	145	148
catalog-or-metadata	53	76
stock-plan-contract	91	23
upstream-early-return	33	35
non-sql-program	17	17
procedure-call	12	4
schema-only	4	6
plan-only	3	4
format-api	1	5
datastream-bypass	2	2
configuration-only	1	1

Provenance¶

These corpora run unchanged release tests with StreamFusion installed. Every passing CSV row joins to one recorded JUnit invocation; skips and failures remain visible in the report but are excluded from passing-case totals. The suite column identifies the selected corpus; runtime includes SQL, Table API, function, batch, catalog and compiled-plan fixtures.

Original CI run.

Both upstream runtime test steps passed, as did all 65 Flink 1.18 and 77 Flink 2.2 native/fallback execution contracts. The Flink 2.2 runtime job is red because the initial CSV exporter rejected an unpaired Unicode surrogate in an upstream fixture; the corrected exporter and classifier were rerun against all original XML and JSON observations. All 11 connector/format jobs passed. Every passing invocation joins exactly once, with no missing or stale evidence. Runtime observations were collected at 8fec58b7195acefa277fd293fbf49ea979706826 and connector observations at baed8451b6fbb7c358b7cd068ab387a4e198d9cb; both use the unchanged engine from b913ce8947582f884356120922ab17f541913d58. The published labels apply the final classifier in this snapshot. The separate state-backend suite and legacy Delta 1.18 host audit are outside this inventory.

Flink 1.18: StreamFusion revisions 8fec58b7195acefa277fd293fbf49ea979706826, baed8451b6fbb7c358b7cd068ab387a4e198d9cb; 6,053 reported cases; outcomes {'passed': 6014, 'skipped': 39}.
Flink 2.2: StreamFusion revisions 8fec58b7195acefa277fd293fbf49ea979706826, baed8451b6fbb7c358b7cd068ab387a4e198d9cb; 9,213 reported cases; outcomes {'passed': 9155, 'skipped': 58}.

CSV exports escape NUL and unpaired Unicode surrogates as \uXXXX; the compressed JSON preserves the exact strings. The JSON also contains observed SQL, original/final plan operators, fallback reasons, planner modes and translation errors. query_verdicts checks each final root for native computation with no remaining host interior operators. Fallback reasons are recorded per optimizer call and can cover several roots; they are not attributed to a specific SQL string.

Native source/sink/format implementation is an additional connector optimization, separate from whole-query SQL admission. Existing native-work contracts remain separate evidence.

See the upstream suite for reproduction, evidence semantics and the separate native-work contracts. Follow-up coverage accounting is tracked in #168.