Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
Impala 3.1.0, Impala 3.2.0, Impala 3.3.0, Impala 3.4.0
-
None
-
ghx-label-14
Description
The LocalCatalog metrics of "CatalogFetch.Partitions.Hits" and "CatalogFetch.Partitions.Requests" in the profile are overcounted. For query "select * from functional.alltypes" where "functional.alltypes" contains 24 partitions, the partition metrics on a cold-started LocalCatalog coordinator are:
- CatalogFetch.Partitions.Hits: 48 - CatalogFetch.Partitions.Misses: 24 - CatalogFetch.Partitions.Requests: 72
Actually, only 48 requests are made on the local cache. 24 of them come from partition pruning and encounter cache miss. Another 24 of them come fromĀ LocalFsTable.toThriftDescriptor() and hit cache.
The overcounting is due to a bug at https://github.com/apache/impala/blob/f8c28f8adfd781727c311b15546a532ce65881e0/fe/src/main/java/org/apache/impala/catalog/local/CatalogdMetaProvider.java#L898
Code snipper:
public Map<String, PartitionMetadata> loadPartitionsByRefs(TableMetaRef table, ...... final int numHits = refToMeta.size(); final int numMisses = partitionRefs.size() - numHits; // Load the remainder from the catalogd. List<PartitionRef> missingRefs = new ArrayList<>(); for (PartitionRef ref: partitionRefs) { if (!refToMeta.containsKey(ref)) missingRefs.add(ref); } if (!missingRefs.isEmpty()) { Map<PartitionRef, PartitionMetadata> fromCatalogd = loadPartitionsFromCatalogd( refImpl, hostIndex, missingRefs); refToMeta.putAll(fromCatalogd); // <---- refToMeta is updated here! // Write back to the cache. storePartitionsInCache(refImpl, hostIndex, fromCatalogd); } sw.stop(); addStatsToProfile(PARTITIONS_STATS_CATEGORY, refToMeta.size(), numMisses, sw); // <--- Should use numHits instead of refToMeta.size() LOG.trace("Request for partitions of {}: hit {}/{}", table, refToMeta.size(), partitionRefs.size());