Generally speaking, as long as a rowset has a version, it can be
considered not to be in a pending state. However, if the rowset was
created through ingesting binlogs, it will have a version but should
still be considered in a pending state because the ingesting txn has not
yet been committed.
This PR updates the condition for determining the pending state. If a
rowset is COMMITTED, the txn should be allowed to roll back even if a
version exists.
Cherry-pick #36551
## Proposed changes
1. return error when bloom filter allocate memory failed
2. return error when deserialize a block， it may need a lot of memory.
---------
Co-authored-by: yiguolei <yiguolei@gmail.com>
## Proposed changes
Issue Number: close #xxx
<!--Describe your changes.-->
## Further comments
If this is a relatively large or complex change, kick off the discussion
at [dev@doris.apache.org](mailto:dev@doris.apache.org) by explaining why
you chose the solution you did and what alternatives you considered,
etc...
* Issue: Doris occasionally encounters an issue where memory usage becomes exceptionally high and does not decrease. The leaked memory is occupied by Bloom filters stored in memory.
Reason: The segment cache stores segment objects read from files into memory. It functions as an LRU cache with an eviction strategy: when the number of segments exceeds the maximum number, or the total memory size of segment objects in the cache exceeds the maximum usage, it evicts the older segments. However, there is a piece of logic in the code that first reads the segment object into memory, assuming it occupies memory size A, then places the read segment object into the cache (at this point, the cache considers the segment object size to be A). It then reads the segment's Bloom filter from the file and assigns it to the segment's Bloom filter member variable, assuming the Bloom filter occupies memory size B. Thus, the total size of the segment object at this point is A+B. However, the cache does not update this size, leading to the actual size of the segment object stored in the cache (A+B) being larger than the size considered by the cache (A). When the number of segment objects in the cache increases to a certain extent, the used memory will surge dramatically. However, the cache does not perceive the size as reaching the eviction limit, so it does not evict the segment objects. In such cases, a memory leak issue arises.
Solution: Since each segment object only reads the Bloom filter once, the issue can be resolved by changing the logic from reading the segment, placing it into the cache, and then reading the Bloom filter to reading the segment, reading the Bloom filter, and then placing it into the cache.
* [fix](compaction)Fix single compaction to get all local versions #33849
add test and comment
* remove single replica compaction prepare input rowsets
reviesd
In order to add common code to the value deleter of LRU cache, let all lru cache values inherit from LRUCacheValueBase class and tracking memory in destructor.
Remove LRUCacheValueBase, put last_visit_time into LRUHandle, and automatically update timestamp to last_visit_time during cache insert and lookup.
Do not rely on external modification of last_visit_time, which is often forgotten.
After all LRU Cache inherits from LRUCachePolicy, this will allow prune stale entry, eviction when memory exceeds limit, and define common properties. LRUCache constructor change to private, only allow LRUCachePolicy to construct it.
Impl DummyLRUCache, when LRU Cache capacity is 0, will no longer be meaningless insert and evict.
1. Do not retry publishing on be When there are too many missing versions, just
add to async publish task.
2. To reduce memory consumption, clean up the tasks when there are too many
async publish tasks.
