messages listlengths 6 6 | sql_fragments listlengths 1 3 | resolution_status stringclasses 1
value | diagnostic_summary stringclasses 3
values | dialogue_transcript stringclasses 3
values |
|---|---|---|---|---|
[
{
"role": "user",
"content": "I am keeping the latest status for each order with this Flink SQL, but the sink rejects the job: INSERT INTO current_orders SELECT order_id, status, update_ts FROM (SELECT order_id, status, update_ts, ROW_NUMBER() OVER (PARTITION BY order_id ORDER BY update_ts DESC) AS rn FROM ... | [
"INSERT INTO current_orders SELECT order_id, status, update_ts FROM (SELECT order_id, status, update_ts, ROW_NUMBER() OVER (PARTITION BY order_id ORDER BY update_ts DESC) AS rn FROM order_updates) ranked WHERE rn = 1"
] | proposed_fix | The latest-row query can emit updates or retractions when a newer order status arrives. An append-only sink cannot represent those changes; use a compatible keyed upsert sink or change the output contract to an append-only event history. | User: I am keeping the latest status for each order with this Flink SQL, but the sink rejects the job: INSERT INTO current_orders SELECT order_id, status, update_ts FROM (SELECT order_id, status, update_ts, ROW_NUMBER() OVER (PARTITION BY order_id ORDER BY update_ts DESC) AS rn FROM order_updates) ranked WHERE rn = 1. ... |
[
{
"role": "user",
"content": "A Flink SQL query groups device readings into five-minute event-time windows, but no rows appear even though the source is receiving data. The query is: SELECT device_id, TUMBLE_START(event_time, INTERVAL '5' MINUTE) AS window_start, COUNT(*) AS reading_count FROM device_readin... | [
"SELECT device_id, TUMBLE_START(event_time, INTERVAL '5' MINUTE) AS window_start, COUNT(*) AS reading_count FROM device_readings GROUP BY device_id, TUMBLE(event_time, INTERVAL '5' MINUTE)",
"WATERMARK FOR event_time AS event_time - INTERVAL '10' SECOND"
] | proposed_fix | The five-minute event-time aggregation appears structurally consistent. An idle source partition may hold back watermark progress and prevent windows from closing; verify partition idleness and watermark advancement before changing window semantics. | User: A Flink SQL query groups device readings into five-minute event-time windows, but no rows appear even though the source is receiving data. The query is: SELECT device_id, TUMBLE_START(event_time, INTERVAL '5' MINUTE) AS window_start, COUNT(*) AS reading_count FROM device_readings GROUP BY device_id, TUMBLE(event_... |
[
{
"role": "user",
"content": "My one-minute aggregation fails with a type error. The query is: SELECT metric_name, TUMBLE_START(event_time_ms, INTERVAL '1' MINUTE) AS window_start, AVG(reading) AS avg_reading FROM readings GROUP BY metric_name, TUMBLE(event_time_ms, INTERVAL '1' MINUTE). I expected an avera... | [
"SELECT metric_name, TUMBLE_START(event_time_ms, INTERVAL '1' MINUTE) AS window_start, AVG(reading) AS avg_reading FROM readings GROUP BY metric_name, TUMBLE(event_time_ms, INTERVAL '1' MINUTE)",
"event_ts AS TO_TIMESTAMP_LTZ(event_time_ms, 3), WATERMARK FOR event_ts AS event_ts - INTERVAL '10' SECOND",
"SELECT... | proposed_fix | The query passes a BIGINT epoch-millisecond field directly to TUMBLE, which requires a suitable timestamp or row-time attribute. Define a computed timestamp with TO_TIMESTAMP_LTZ, declare a watermark on it, and use it in the window expressions. | User: My one-minute aggregation fails with a type error. The query is: SELECT metric_name, TUMBLE_START(event_time_ms, INTERVAL '1' MINUTE) AS window_start, AVG(reading) AS avg_reading FROM readings GROUP BY metric_name, TUMBLE(event_time_ms, INTERVAL '1' MINUTE). I expected an average per metric per minute. || Assista... |
Streaming Query and Transformation Debugging Dialogue Dataset
This dataset contains multi-turn conversations in which users progressively provide streaming SQL, transformation logic, error messages, and expected results, along with query fragments and assistant responses. It focuses on diagnosing syntax and logic issues in real-time data pipeline development, capturing an iterative workflow from clarifying context to proposing revisions. It is suited to dialogue supervised fine-tuning, streaming development assistant training, and evaluation of SQL debugging and query repair capabilities.
Technical Specifications
| Field | Type | Description |
|---|---|---|
| messages | array | The transcript parsed into an ordered sequence of user and assistant messages. |
| sql_fragments | array | SQL snippets extracted from the dialogue for debugging, review, or revision. |
| resolution_status | string | The outcome reflected in the dialogue, such as a proposed fix or an unresolved issue. |
| diagnostic_summary | string | A concise summary of the suspected syntax or logic issue and its relevant context. |
| dialogue_transcript | string | The complete conversation, including user-provided SQL, transformation logic, errors, expected results, and assistant responses. |
Compliance Statement
| Authorization Type | CC-BY-NC-SA 4.0 (Attribution–NonCommercial–ShareAlike) |
| Commercial Use | Requires exclusive subscription or authorization contract (monthly or per-invocation charging) |
| Privacy and Anonymization | No PII, no real company names, simulated scenarios follow industry standards |
| Compliance System | Compliant with China's Data Security Law / EU GDPR / supports enterprise data access logs |
Source & Contact
- Downloads last month
- -