RFile.Writer.append checks for giant keys and avoid writing them as index blocks. This check is flawed and can result in multi-GB blocks. In our case, a 20GB compressed RFile had one block with over 2GB raw size. This happened because the key size statistics changed after some point in the file. The code in question follows:
Before closing a block that has grown beyond the target block size we check to see that the key is below average in size or that the block is 1.1 times the target block size (maxBlockSize), and we check that the key isn't a "giant" key, or more than 3 standard deviations from the mean of keys seen so far.
Our RFiles often have one row of data with different column families representing various forward and inverted indexes. This is a table design similar to the WikiSearch example. The first column family in this case had very uniform, relatively small key sizes. This first column family comprised gigabytes of data, split up into roughly 100KB blocks. When we switched to the next column family the keys grew in size, but were still under about 100 bytes. The statistics of the first column family had firmly established a smaller mean and tiny standard deviation (approximately 0), and it took over 2GB of larger keys to bring the standard deviation up enough so that keys were no longer considered "giant" and the block could be closed.
Now that we're aware, we see large blocks (more than 10x the target block size) in almost every RFile we write. This only became a glaring problem when we got OOM exceptions trying to decompress the block, but it also shows up in a number of subtle performance problems, like high variance in latencies for looking up particular keys.
The fix for this should produce bounded RFile block sizes, limited to the greater of 2x the maximum key/value size in the block and some configurable threshold, such as 1.1 times the compressed block size. We need a firm cap to be able to reason about memory usage in various applications.
The following code produces arbitrarily large RFile blocks:
One telltale symptom of this bug is an OutOfMemoryException thrown from a readahead thread with message "Requested array size exceeds VM limit". This will only happen if the block cache size is big enough to hold the expected raw block size, 2GB in our case. This message is rare, and really only happens when allocating an array of size Integer.MAX_VALUE or Integer.MAX_VALUE-1 on the hotspot JVM. Integer.MAX_VALUE happens in this case due to some strange handling of raw block sizes in the BCFile code. Most OutOfMemoryExceptions have different messages.