[IMPALA-8690] Better eviction algorithm for data cache

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: Impala 3.3.0
• Fix Version/s: Impala 4.0.0
• Component/s: Backend
• Epic Color: ghx-label-9
• Assignee: Joe McDonnell
• Reporter: Michael Ho
• Created: 20/Jun/19 20:18
• Updated: 02/Jun/20 17:27
• Resolved: 04/Apr/20 00:12

Description

With the current implementation of data cache, all data access will be cached regardless of the access pattern. The current LRU eviction algorithm is not resistant to scan traffic so in case some users scan a big fact table, a lot of the heavily accessed items will be evicted inevitably. We should adopt better eviction algorithm (e.g. LRFU or some other well known ones in the literature). Would be nice to evaluate it against some users' traces now that IMPALA-8542 is fixed.

In the short run, we probably need some workaround (e.g. query hints to disable caching for certain tables). Will file a separate jira for it.

Issue Links

• is related to: IMPALA-8691 Query hint for disabling data caching
• relates to: IMPALA-8341 Data cache for remote reads