[HBASE-17874] Limiting of read request response size based on block size may go wrong when blocks are read from onheap or off heap bucket cache

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Invalid
• Affects Version/s: 2.0.0
• Fix Version/s: 2.0.0
• Assignee: Anoop Sam John
• Reporter: Anoop Sam John
• Created: 04/Apr/17 16:34
• Updated: 01/Aug/18 06:22
• Resolved: 26/May/17 12:47

Description

HBASE-14978 added this size limiting so as to make sure the multi read requests do not retain two many blocks. This works well when the blocks are obtained from any where other than memory mode BucketCache. In case of on heap or off heap Bucket Cache, the entire cache area is split into N ByteBuffers each of size 4 MB. When we hit a block in this cache, we no longer do copy data into temp array. We use the same shared memory (BB). Its capacity is 4 MB.
The block size accounting logic is RSRpcServices is like below

 if (c instanceof ByteBufferCell) {
          ByteBufferCell bbCell = (ByteBufferCell) c;
          ByteBuffer bb = bbCell.getValueByteBuffer();
          if (bb != lastBlock) {
            context.incrementResponseBlockSize(bb.capacity());
            lastBlock = bb;
          }
        } else {
          // We're using the last block being the same as the current block as
          // a proxy for pointing to a new block. This won't be exact.
          // If there are multiple gets that bounce back and forth
          // Then it's possible that this will over count the size of
          // referenced blocks. However it's better to over count and
          // use two rpcs than to OOME the regionserver.
          byte[] valueArray = c.getValueArray();
          if (valueArray != lastBlock) {
            context.incrementResponseBlockSize(valueArray.length);
            lastBlock = valueArray;
          }
        }
		

We take the BBCell's value buffer and takes its capacity. The cell is backed by the same BB that backs the HFileBlock. When the HFileBlock is created from the BC, we do as below duplicating and proper positioning and limiting the BB

 ByteBuffer bb = buffers[i].duplicate();
      if (i == startBuffer) {
        cnt = bufferSize - startBufferOffset;
        if (cnt > len) cnt = len;
        bb.limit(startBufferOffset + cnt).position(startBufferOffset);
		

Still this BB's capacity is 4 MB.
This will make the size limit breach to happen too soon. What we expect is block size defaults to 64 KB and so we here by allow cells from different blocks to appear in response. We have a way to check whether we move from one block to next.

if (bb != lastBlock) {
...
            lastBlock = bb;
}

But already just by considering the 1st cell, we added 4 MB size!

Issue Links

• is related to: HBASE-14978 Don't allow Multi to retain too many blocks