[
https://issues.apache.org/jira/browse/CASSANDRA-13304?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15926927#comment-15926927
]
Michael Kjellman commented on CASSANDRA-13304:
----------------------------------------------
[~mpenick] I personally have never seen Cassandra come close to saturating a
NIC in all my years of production experience... the best place where we might
be able to do so would be in streaming but due to the current streaming design
we have a billion other bottlenecks before we hit the NIC... what I've seen is
that NIC offloading is enabled by default on most NICs (even onboard
motherboard ones), so yes, potentially if we wanted to say "hey -- we will be
protected by ECC in main RAM and all our boxes are configured properly to
kernel panic when an ECC uncorrectable is hit -- and we've got TCP offloading
disabled so all our memory is protected" -- we still could have bits flipped in
in one of the other components i wouldn't even be able to think could fail,
until it does..
Regardless, even if the NIC became a bottleneck, the correct tradeoff should
still be to ensure data integrity (given we are a database) even if that means
we lower actual possible throughput.. everything always has a tradeoff...
As an additional aside: with JDK8, CRC32 calculations are now using the x86
hardware instruction so the digest calculations are really quick
(https://bugs.openjdk.java.net/browse/JDK-7088419). There is similar work in
JDK9 to make adler32 computations use the hardware instruction set too.
> Add checksumming to the native protocol
> ---------------------------------------
>
> Key: CASSANDRA-13304
> URL: https://issues.apache.org/jira/browse/CASSANDRA-13304
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Michael Kjellman
> Assignee: Michael Kjellman
> Labels: client-impacting
> Attachments: 13304_v1.diff
>
>
> The native binary transport implementation doesn't include checksums. This
> makes it highly susceptible to silently inserting corrupted data either due
> to hardware issues causing bit flips on the sender/client side, C*/receiver
> side, or network in between.
> Attaching an implementation that makes checksum'ing mandatory (assuming both
> client and server know about a protocol version that supports checksums) --
> and also adds checksumming to clients that request compression.
> The serialized format looks something like this:
> {noformat}
> * 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3
> * 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Number of Compressed Chunks | Compressed Length (e1) /
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * / Compressed Length cont. (e1) | Uncompressed Length (e1) /
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Uncompressed Length cont. (e1)| CRC32 Checksum of Lengths (e1)|
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Checksum of Lengths cont. (e1)| Compressed Bytes (e1) +//
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | CRC32 Checksum (e1) ||
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Compressed Length (e2) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Uncompressed Length (e2) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | CRC32 Checksum of Lengths (e2) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Compressed Bytes (e2) +//
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | CRC32 Checksum (e2) ||
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Compressed Length (en) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Uncompressed Length (en) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | CRC32 Checksum of Lengths (en) |
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | Compressed Bytes (en) +//
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> * | CRC32 Checksum (en) ||
> * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> {noformat}
> The first pass here adds checksums only to the actual contents of the frame
> body itself (and doesn't actually checksum lengths and headers). While it
> would be great to fully add checksuming across the entire protocol, the
> proposed implementation will ensure we at least catch corrupted data and
> likely protect ourselves pretty well anyways.
> I didn't go to the trouble of implementing a Snappy Checksum'ed Compressor
> implementation as it's been deprecated for a while -- is really slow and
> crappy compared to LZ4 -- and we should do everything in our power to make
> sure no one in the community is still using it. I left it in (for obvious
> backwards compatibility aspects) old for clients that don't know about the
> new protocol.
> The current protocol has a 256MB (max) frame body -- where the serialized
> contents are simply written in to the frame body.
> If the client sends a compression option in the startup, we will install a
> FrameCompressor inline. Unfortunately, we went with a decision to treat the
> frame body separately from the header bits etc in a given message. So,
> instead we put a compressor implementation in the options and then if it's
> not null, we push the serialized bytes for the frame body *only* thru the
> given FrameCompressor implementation. The existing implementations simply
> provide all the bytes for the frame body in one go to the compressor
> implementation and then serialize it with the length of the compressed bytes
> up front.
> Unfortunately, this won't work for checksum'ing for obvious reasons as we
> can't naively just checksum the entire (potentially) 256MB frame body and
> slap it at the end... so,
> The best place to start with the changes is in {{ChecksumedCompressor}}. I
> implemented one single place to perform the checksuming (and to support
> checksuming) the actual required chunking logic. Implementations of
> ChecksumedCompressor only implement the actual calls to the given compression
> algorithm for the provided bytes.
> Although the interface takes a {{Checksum}}, right now the attached patch
> uses CRC32 everywhere. As of right now, given JDK8+ has support for doing the
> calculation with the Intel instruction set, CRC32 is about as fast as we can
> get right now.
> I went with a 32kb "default" for the chunk size -- meaning we will chunk the
> entire frame body into 32kb chunks, compress each one of those chunks, and
> checksum the chunk. Upon discussing with a bunch of people and researching
> how checksums actually work and how much data they will protect etc -- if we
> use 32kb chunks with CRC32 we can catch up to 32 bits flipped in a row (but
> more importantly catch the more likely corruption where a single bit is
> flipped) with pretty high certainty. 64kb seems to introduce too much of a
> probability of missing corruption.
> The maximum block size LZ4 operates on is a 64kb chunk -- so this combined
> with the need to make sure the CRC32 checksums are actually going to catch
> stuff -- chunking at 32kb seemed like a good reasonable value to use when
> weighing both checksums and compression (to ensure we don't kill our
> compression ratio etc).
> I'm not including client changes here -- I asked around and I'm not really
> sure what the policy there is -- do we update the python driver? java driver?
> how has the timing of this stuff been handled in the past?
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
