Change-Id: I28f7800030a3b9db48e315061509af0746feffcc
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>

By showing the bootstrap layer in getAdditionalRefs() garbage
collector code can be more RefDatabase agnostic and not care about
the special case of RefTree and RefTreeNames for the purposes of
building up the roots to GC. Instead they can combine getRefs(ALL)
and getAdditionalRefs() and have a clean set of roots.

Change-Id: I665cd2456e9316640215b6a08bc728d1356f36d8

The RefTree graph needs to be quickly accessed to read references.
It is also distinct graph disconnected from the rest of the
repository. Store the commit and tree objects in their own pack.

Change-Id: Icbb735be8fa91ccbf0708ca3a219b364e11a6b83

Instead of storing references in the local filesystem rely on the
RefTree rooted at refs/txn/committed.  This avoids needing to store
references in the packed-refs file by keeping all data rooted under
a single refs/txn/committed ref.

Performance to scan all references from a well packed RefTree is very
close to reading the packed-refs file from local disk.

Storing a packed RefTree is smaller due to pack file compression,
about 49.39 bytes/ref (on average) compared to packed-refs using
~65.49 bytes/ref.

Change-Id: I75caa631162dc127a780095066195cbacc746d49

Change-Id: If8522d4e02254766e3310a70e42badafa5ab9768
Signed-off-by: Robin Rosenberg <robin.rosenberg@dewire.com>

This breaks all existing callers once. Applications are not supposed
to build against the internal storage API unless they can accept API
churn and make necessary updates as versions change.

Change-Id: I2ab1327c202ef2003565e1b0770a583970e432e9

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

jgit.storage.dht is a storage provider implementation for JGit that
permits storing the Git repository in a distributed hashtable, NoSQL
system, or other database.  The actual underlying storage system is
undefined, and can be plugged in by implementing 7 small interfaces:

  *  Database
  *  RepositoryIndexTable
  *  RepositoryTable
  *  RefTable
  *  ChunkTable
  *  ObjectIndexTable
  *  WriteBuffer

The storage provider interface tries to assume very little about the
underlying storage system, and requires only three key features:

  *  key -> value lookup (a hashtable is suitable)
  *  atomic updates on single rows
  *  asynchronous operations (Java's ExecutorService is easy to use)

Most NoSQL database products offer all 3 of these features in their
clients, and so does any decent network based cache system like the
open source memcache product.  Relying only on key equality for data
retrevial makes it simple for the storage engine to distribute across
multiple machines.  Traditional SQL systems could also be used with a
JDBC based spi implementation.

Before submitting this change I have implemented six storage systems
for the spi layer:

  * Apache HBase[1]
  * Apache Cassandra[2]
  * Google Bigtable[3]
  * an in-memory implementation for unit testing
  * a JDBC implementation for SQL
  * a generic cache provider that can ride on top of memcache

All six systems came in with an spi layer around 1000 lines of code to
implement the above 7 interfaces.  This is a huge reduction in size
compared to prior attempts to implement a new JGit storage layer.  As
this package shows, a complete JGit storage implementation is more
than 17,000 lines of fairly complex code.

A simple cache is provided in storage.dht.spi.cache.  Implementers can
use CacheDatabase to wrap any other type of Database and perform fast
reads against a network based cache service, such as the open source
memcached[4].  An implementation of CacheService must be provided to
glue this spi onto the network cache.

[1] https://github.com/spearce/jgit_hbase
[2] https://github.com/spearce/jgit_cassandra
[3] http://labs.google.com/papers/bigtable.html
[4] http://memcached.org/

Change-Id: I0aa4072781f5ccc019ca421c036adff2c40c4295
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

There is no point in pushing all of the files within the edge
commits into the delta search when making a thin pack.  This floods
the delta search window with objects that are unlikely to be useful
bases for the objects that will be written out, resulting in lower
data compression and higher transfer sizes.

Instead observe the path of a tree or blob that is being pushed
into the outgoing set, and use that path to locate up to WINDOW
ancestor versions from the edge commits.  Push only those objects
into the edgeObjects set, reducing the number of objects seen by the
search window.  This allows PackWriter to only look at ancestors
for the modified files, rather than all files in the project.
Limiting the search to WINDOW size makes sense, because more than
WINDOW edge objects will just skip through the window search as
none of them need to be delta compressed.

To further improve compression, sort edge objects into the front
of the window list, rather than randomly throughout.  This puts
non-edges later in the window and gives them a better chance at
finding their base, since they search backwards through the window.

These changes make a significant difference in the thin-pack:

  Before:
    remote: Counting objects: 144190, done
    remote: Finding sources: 100% (50275/50275)
    remote: Getting sizes: 100% (101405/101405)
    remote: Compressing objects: 100% (7587/7587)
    Receiving objects: 100% (50275/50275), 24.67 MiB | 9.90 MiB/s, done.
    Resolving deltas: 100% (40339/40339), completed with 2218 local objects.

    real    0m30.267s

  After:
    remote: Counting objects: 61549, done
    remote: Finding sources: 100% (50275/50275)
    remote: Getting sizes: 100% (18862/18862)
    remote: Compressing objects: 100% (7588/7588)
    Receiving objects: 100% (50275/50275), 11.04 MiB | 3.51 MiB/s, done.
    Resolving deltas: 100% (43160/43160), completed with 5014 local objects.

    real    0m22.170s

The resulting pack is 13.63 MiB smaller, even though it contains the
same exact objects.  82,543 fewer objects had to have their sizes
looked up, which saved about 8s of server CPU time.  2,796 more
objects from the client were used as part of the base object set,
which contributed to the smaller transfer size.

Change-Id: Id01271950432c6960897495b09deab70e33993a9
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Sigend-off-by: Chris Aniszczyk <caniszczyk@gmail.com>

Per CQ 3448 this is the initial contribution of the JGit project
to eclipse.org.  It is derived from the historical JGit repository
at commit 3a2dd9921c.

Signed-off-by: Shawn O. Pearce <spearce@spearce.org>