[HIVE-19261] Avro SerDe's InstanceCache should not be synchronized on retrieve

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-1
• Labels: pull-request-available
• Assignee: Alexey Diomin
• Reporter: Fangshi Li
• Created: 21/Apr/18 07:39
• Updated: 17/Nov/22 08:55
• Resolved: 19/Jun/20 13:36

Description

In HIVE-16175, upstream made a patch to fix the thread safety issue in AvroSerDe's InstanceCache. This fix made the retrieve method in InstanceCache synchronized. While it should make InstanceCache thread-safe, making retrieve synchronized for the cache can be expensive in highly concurrent environment like Spark, as multiple threads need to be synchronized on entering the entire retrieve method.

We are proposing another way to fix this thread safety issue by making the underlying map of InstanceCache as ConcurrentHashMap. Ideally, we can use atomic computeIfAbsent in the retrieve method to avoid synchronizing the entire method.

While computeIfAbsent is only available on java 8 and java 7 is still supported in Hive,
we use a pattern to simulate the behavior of computeIfAbsent. In the future, we should move to computeIfAbsent when Hive requires java 8.

Issue Links

• relates to: HIVE-16175 Possible race condition in InstanceCache
• relates to: HIVE-23731 Review of AvroInstance Cache
• links to: GitHub Pull Request #807