carp84 commented on a change in pull request #10820:
[FLINK-15512][statebackend] Refactor the mechanism to calculate the cache
capacity shared among RocksDB instance(s)
URL: https://github.com/apache/flink/pull/10820#discussion_r365127569
##########
File path:
flink-state-backends/flink-statebackend-rocksdb/src/main/java/org/apache/flink/contrib/streaming/state/RocksDBOperationUtils.java
##########
@@ -213,4 +211,40 @@ public static void addColumnFamilyOptionsToCloseLater(
throw new IOException("Failed to acquire shared cache
resource for RocksDB", e);
}
}
+
+ @VisibleForTesting
+ static RocksDBSharedResources allocateRocksDBSharedResources(long size,
double writeBufferRatio, double highPriorityPoolRatio) {
+ long calculatedCacheSize = calculateActualCacheSize(size,
writeBufferRatio);
+ final Cache cache = createCache(calculatedCacheSize,
highPriorityPoolRatio);
+ final WriteBufferManager wbm = new WriteBufferManager((long)
(writeBufferRatio * size), cache);
+ return new RocksDBSharedResources(cache, wbm);
+ }
+
+ /**
+ * Calculate the actual calculated memory size of cache, which would be
shared among rocksDB instance(s).
+ * We introduce this method because:
+ * a) We cannot create a strict capacity limit cache util FLINK-15532
resolved.
+ * b) Regardless of the memory usage of blocks pinned by RocksDB
iterators,
+ * which is difficult to calculate and only happened when we iterator
entries in RocksDBMapState, the overuse of memory is mainly occupied by at most
half of the write buffer usage.
+ * (see <a
href="https://github.com/dataArtisans/frocksdb/blob/958f191d3f7276ae59b270f9db8390034d549ee0/include/rocksdb/write_buffer_manager.h#L51";>the
flush implementation of write buffer manager</a>).
+ * Thus, we have four equations below:
+ * write_buffer_manager_memory = 1.5 * write_buffer_manager_capacity
+ * write_buffer_manager_memory = total_memory_size *
write_buffer_ratio
+ * write_buffer_manager_memory + other_part = total_memory_size
+ * write_buffer_manager_capacity + other_part = cache_size
+ * And we would deduce the formula: cache_size = 3 * total_memory_size
/ (3 + write_buffer_ratio)
+ *
+ * @param totalMemorySize Total off-heap memory size reserved for
RocksDB instance(s).
+ * @param writeBufferRatio The ratio of memory size which could be
reserved for write buffer manager to control the memory usage.
+ * @return The actual calculated memory size.
+ */
+ @VisibleForTesting
+ static long calculateActualCacheSize(long totalMemorySize, double
writeBufferRatio) {
+ return (long) (3 * totalMemorySize / (3 + writeBufferRatio));
+ }
+
+ @VisibleForTesting
+ static Cache createCache(long cacheSize, double highPriorityPoolRatio) {
+ return new LRUCache(cacheSize, -1, false,
highPriorityPoolRatio);
Review comment:
Suggest to add a TODO comment here, mentioning that we will change the
`strictCapacityLimit` flag to `true` after
[rocksdb#6247](https://github.com/facebook/rocksdb/issues/6247) is resolved.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
