[IMPALA-7178] Reduce logging for common data errors

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: Backend
• Labels: supportability
• Epic Link: SupportabilityRoadMap
• Epic Color: ghx-label-7
• Assignee: Csaba Ringhofer
• Reporter: Csaba Ringhofer
• Created: 15/Jun/18 13:31
• Updated: 07/Jul/20 13:20

Description

Some data errors (for example out-of-range parquet timestamps) can dominate logs if a table contains a large number of rows with invalid data. If an error has its own error code (see common/thrift/generate_error_codes.py), then these errors are already aggregated to the user (RuntimeState::LogError()) for every query, but the logs will contain a new line for every occurrence. This is not too useful most of times, as the log lines will repeat the same information (the corrupt data itself is not logged as it can be sensitive information).

The best would to reduce logging without loosing information:

• the first occurrence of an error should be logged (per query/fragment/table/file/column) to help the investigation of cases where the data error leads to other errors and to avoid breaking log analyzer tools that search for the current format
• other occurrences can be aggregated, like "in query Q table T column C XY error occurred N times"

An extra goal is to avoid calling RuntimeState::LogError() for other occurrences than the first one, as RuntimeState::LogError() uses a (per fragment) lock.

Issue Links

• is related to: IMPALA-5845 Impala should de-duplicate row parsing error