In practice the DHT storage layer has not been performing as well as
large scale server environments want to see from a Git server.
The performance of the DHT schema degrades rapidly as small changes
are pushed into the repository due to the chunk size being less than
1/3 of the pushed pack size. Small chunks cause poor prefetch
performance during reading, and require significantly longer prefetch
lists inside of the chunk meta field to work around the small size.
The DHT code is very complex (>17,000 lines of code) and is very
sensitive to the underlying database round-trip time, as well as the
way objects were written into the pack stream that was chunked and
stored on the database. A poor pack layout (from any version of C Git
prior to Junio reworking it) can cause the DHT code to be unable to
enumerate the objects of the linux-2.6 repository in a completable
time scale.
Performing a clone from a DHT stored repository of 2 million objects
takes 2 million row lookups in the DHT to locate the OBJECT_INDEX row
for each object being cloned. This is very difficult for some DHTs to
scale, even at 5000 rows/second the lookup stage alone takes 6 minutes
(on local filesystem, this is almost too fast to bother measuring).
Some servers like Apache Cassandra just fall over and cannot complete
the 2 million lookups in rapid fire.
On a ~400 MiB repository, the DHT schema has an extra 25 MiB of
redundant data that gets downloaded to the JGit process, and that is
before you consider the cost of the OBJECT_INDEX table also being
fully loaded, which is at least 223 MiB of data for the linux kernel
repository. In the DHT schema answering a `git clone` of the ~400 MiB
linux kernel needs to load 248 MiB of "index" data from the DHT, in
addition to the ~400 MiB of pack data that gets sent to the client.
This is 193 MiB more data to be accessed than the native filesystem
format, but it needs to come over a much smaller pipe (local Ethernet
typically) than the local SATA disk drive.
I also never got around to writing the "repack" support for the DHT
schema, as it turns out to be fairly complex to safely repack data in
the repository while also trying to minimize the amount of changes
made to the database, due to very common limitations on database
mutation rates..
This new DFS storage layer fixes a lot of those issues by taking the
simple approach for storing relatively standard Git pack and index
files on an abstract filesystem. Packs are accessed by an in-process
buffer cache, similar to the WindowCache used by the local filesystem
storage layer. Unlike the local file IO, there are some assumptions
that the storage system has relatively high latency and no concept of
"file handles". Instead it looks at the file more like HTTP byte range
requests, where a read channel is a simply a thunk to trigger a read
request over the network.
The DFS code in this change is still abstract, it does not store on
any particular filesystem, but is fairly well suited to the Amazon S3
or Apache Hadoop HDFS. Storing packs directly on HDFS rather than
HBase removes a layer of abstraction, as most HBase row reads turn
into an HDFS read.
Most of the DFS code in this change was blatently copied from the
local filesystem code. Most parts should be refactored to be shared
between the two storage systems, but right now I am hesistent to do
this due to how well tuned the local filesystem code currently is.
Change-Id: Iec524abdf172e9ec5485d6c88ca6512cd8a6eafb
fa4cc2475f