Revision 1298641

[jira] [HBASE-5074] Support checksums in HBase block cache

Author: Dhruba

Summary:
HFile is enhanced to store a checksum for each block. HDFS checksum verification
is avoided while reading data into the block cache. On a checksum verification
failure, we retry the file system read request with hdfs checksums switched on
(thanks Todd).

I have a benchmark that shows that it reduces iops on the disk by about 40%. In
this experiment, the entire memory on the regionserver is allocated to the
regionserver's jvm and the OS buffer cache size is negligible. I also measured
negligible (<5%) additional cpu usage while using hbase-level checksums.

The salient points of this patch:

1. Each hfile's trailer used to have a 4 byte version number. I enhanced this so
that these 4 bytes can be interpreted as a (major version number, minor
version). Pre-existing hfiles have a minor version of 0. The new hfile format
has a minor version of 1 (thanks Mikhail). The hfile major version remains
unchanged at 2. The reason I did not introduce a new major version number is
because the code changes needed to store/read checksums do not differ much from
existing V2 writers/readers.

2. Introduced a HFileSystem object which is a encapsulates the FileSystem
objects needed to access data from hfiles and hlogs.  HDFS FileSystem objects
already had the ability to switch off checksum verifications for reads.

3. The majority of the code changes are located in hbase.io.hfie package. The
retry of a read on an initial checksum failure occurs inside the hbase.io.hfile
package itself.  The code changes to hbase.regionserver package are minor.

4. The format of a hfileblock is the header followed by the data followed by the
checksum(s). Each 16 K (configurable) size of data has a 4 byte checksum.  The
hfileblock header has two additional fields: a 4 byte value to store the
bytesPerChecksum and a 4 byte value to store the size of the user data
(excluding the checksum data). This is well explained in the associated
javadocs.

5. I added a test to test backward compatibility. I will be writing more unit
tests that triggers checksum verification failures aggressively. I have left a
few redundant log messages in the code (just for easier debugging) and will
remove them in later stage of this patch. I will also be adding metrics on
number of checksum verification failures/success in a later version of this
diff.

6. By default, hbase-level checksums are switched on and hdfs level checksums
are switched off for hfile-reads. No changes to Hlog code path here.

Test Plan: The default setting is to switch on hbase checksums for hfile-reads,
thus all existing tests actually validate the new code pieces. I will be writing
more unit tests for triggering checksum verification failures.

Reviewers: mbautin

Reviewed By: mbautin

CC: JIRA, tedyu, mbautin, dhruba, todd, stack

Differential Revision: https://reviews.facebook.net/D1521