[IMPALA-8937] Fine grained table metadata loading on Catalog server

• Type: Epic
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: Impala 2.12.0, Impala 3.3.0
• Component/s: Catalog, Frontend
• Labels: 2023Q1
• Epic Name: fine-grained metadata in catalogd
• Epic Color: ghx-label-11
• Assignee: Unassigned
• Reporter: Bharath Vissapragada
• Created: 11/Sep/19 04:30
• Updated: 21/Aug/23 09:10

Description

Background:

Currently the table on the Catalog server is either in a loaded or unloaded state (IncompleteTable). When Catalog server starts for the first time, we first fetch a list of table names for each databases and every table in this list starts as an unloaded table. The table lists are propagated to the coordinators so that they know whether a table with a given name exists or not and they can start analyzing the queries. No metadata is loaded in the incomplete tables (like schema/ownership, comments etc.)

The table metadata is loaded lazily (and the table moves into a loaded state) when it is referenced in any query. When a load request comes in, all the table metadata is loaded including file block information. 

Problem: 

Coordinators need some additional information when analyzing unloaded tables. For example: IMPALA-8228. The ownership information is a part of the HMS table schema which is not loaded until the table is marked fully loaded. While this is not a problem for regular queries (like select * from <tbl>), it is an issue with queries like "show tables" which do not trigger a table load. In this particular case, due to the lack of ownership information, the output of the table listing could be different depending on whether the table is loaded. Another example is IMPALA-8606 where the GET_TABLES request does not return the table comments because they are not available for unloaded tables.

Ask:

We need to consider finer grained loading on the Catalog server in general. Instead of having a binary state (loaded vs unloaded), the table could be in a partially loaded state. We could also start with aggressively fetching certain pieces of information that we think could aid with analysis and lazily load the remaining pieces of metadata. Finer grained loading also integrates well with the LocalCatalog implementation on the coordinators where the the entire table need not be loaded on the Catalog server to serve partial meta information (e.g: show partitions <large-table>).

Issue Links

• is related to: IMPALA-7501 Slim down metastore Partition objects in LocalCatalog cache
• is related to: IMPALA-7533 Optimize fetch-from-catalog by caching partitions across table versions
• is related to: IMPALA-9936 Only send invalidations in DDL responses to LocalCatalog coordinators
• is related to: IMPALA-5468 REFRESH and a very large table run after catalog restart cannot be cancelled
• is related to: IMPALA-8228 Support for object ownership with Ranger authorization provider
• is related to: IMPALA-6354 Consider using Guava LoadingCache to cache metadata objects opposed to a ConcurrentHashMap
• is related to: IMPALA-8953 Tables and Databases sharing same name can cause query failures if table is not readable by Impala