#LyX 1.4.2 created this file. For more info see http://www.lyx.org/
\lyxformat 245
\begin_document
\begin_header
\textclass article
\language english
\inputencoding auto
\fontscheme default
\graphics default
\paperfontsize default
\spacing single
\papersize default
\use_geometry false
\use_amsmath 1
\cite_engine basic
\use_bibtopic false
\paperorientation portrait
\secnumdepth 3
\tocdepth 3
\paragraph_separation skip
\defskip medskip
\quotes_language english
\papercolumns 1
\papersides 1
\paperpagestyle default
\tracking_changes false
\output_changes false
\end_header

\begin_body

\begin_layout Title

\size larger
Automatic File Replication (replicate) in GlusterFS
\end_layout

\begin_layout Author
Vikas Gorur 
\family typewriter
\size larger
<vikas@gluster.com>
\end_layout

\begin_layout Standard
\begin_inset ERT
status open

\begin_layout Standard


\backslash
hrule
\end_layout

\end_inset


\end_layout

\begin_layout Section*
Overview
\end_layout

\begin_layout Standard
This document describes the design and usage of the replicate translator in GlusterFS.
 This document is valid for the 1.4.x releases, and not earlier ones.
\end_layout

\begin_layout Standard
The replicate translator of GlusterFS aims to keep identical copies of a file
 on all its subvolumes, as far as possible.
 It tries to do this by performing all filesystem mutation operations (writing
 data, creating files, changing ownership, etc.) on all its subvolumes in
 such a way that if an operation succeeds on atleast one subvolume, all
 other subvolumes can later be brought up to date.
\end_layout

\begin_layout Standard
In the rest of the document the terms 
\begin_inset Quotes eld
\end_inset

subvolume
\begin_inset Quotes erd
\end_inset

 and 
\begin_inset Quotes eld
\end_inset

server
\begin_inset Quotes erd
\end_inset

 are used interchangeably, trusting that it will cause no confusion to the
 reader.
\end_layout

\begin_layout Section*
Usage
\end_layout

\begin_layout Standard
A sample volume declaration for replicate looks like this:
\end_layout

\begin_layout Standard
\begin_inset ERT
status open

\begin_layout Standard


\backslash
begin{verbatim}
\end_layout

\begin_layout Standard

volume replicate
\end_layout

\begin_layout Standard

 type cluster/replicate
\end_layout

\begin_layout Standard

 # options, see below for description
\end_layout

\begin_layout Standard

 subvolumes brick1 brick2
\end_layout

\begin_layout Standard

end-volume
\end_layout

\begin_layout Standard


\backslash
end{verbatim}
\end_layout

\begin_layout Standard

\end_layout

\begin_layout Standard

\end_layout

\begin_layout Standard

\end_layout

\end_inset


\end_layout

\begin_layout Standard
This defines an replicate volume with two subvolumes, brick1, and brick2.
 For replicate to work properly, it is essential that its subvolumes support 
\series bold
extended attributes
\series default
.
 This means that you should choose a backend filesystem that supports extended
 attributes, like XFS, ReiserFS, or Ext3.
\end_layout

\begin_layout Standard
The storage volumes used as backend for replicate 
\emph on
must
\emph default
 have a posix-locks volume loaded above them.
\end_layout

\begin_layout Standard
\begin_inset ERT
status open

\begin_layout Standard


\backslash
begin{verbatim}
\end_layout

\begin_layout Standard

volume brick1
\end_layout

\begin_layout Standard

 type features/posix-locks
\end_layout

\begin_layout Standard

 subvolumes brick1-ds
\end_layout

\begin_layout Standard

end-volume
\end_layout

\begin_layout Standard


\backslash
end{verbatim}
\end_layout

\end_inset


\end_layout

\begin_layout Section*
Design
\end_layout

\begin_layout Subsection*
Read algorithm
\end_layout

\begin_layout Standard
All operations that do not modify the file or directory are sent to all
 the subvolumes and the first successful reply is returned to the application.
\end_layout

\begin_layout Standard
The read() system call (reading data from a file) is an exception.
 For read() calls, replicate tries to do load balancing by sending all reads from
 a particular file to a particular server.
\end_layout

\begin_layout Standard
The read algorithm is also affected by the option read-subvolume; see below
 for details.
\end_layout

\begin_layout Subsection*
Classes of file operations
\end_layout

\begin_layout Standard
replicate divides all filesystem write operations into three classes:
\end_layout

\begin_layout Itemize

\series bold
data: 
\series default
Operations that modify the contents of a file (write, truncate).
\end_layout

\begin_layout Itemize

\series bold
metadata: 
\series default
Operations that modify attributes of a file or directory (permissions, ownership
, etc.).
\end_layout

\begin_layout Itemize

\series bold
entry: 
\series default
Operations that create or delete directory entries (mkdir, create, rename,
 rmdir, unlink, etc.).
\end_layout

\begin_layout Subsection*
Locking and Change Log
\end_layout

\begin_layout Standard
To ensure consistency across subvolumes, replicate holds a lock whenever a modificatio
n is being made to a file or directory.
 By default, replicate considers the first subvolume as the sole lock server.
 However, the number of lock servers can be increased upto the total number
 of subvolumes.
\end_layout

\begin_layout Standard
The change log is a set of extended attributes associated with files and
 directories that replicate maintains.
 The change log keeps track of the changes made to files and directories
 (data, metadata, entry) so that the self-heal algorithm knows which copy
 of a file or directory is the most recent one.
\end_layout

\begin_layout Subsection*
Write algorithm
\end_layout

\begin_layout Standard
The algorithm for all write operations (data, metadata, entry) is:
\end_layout

\begin_layout Enumerate
Lock the file (or directory) on all of the lock servers (see options below).
\end_layout

\begin_layout Enumerate
Write change log entries on all servers.
\end_layout

\begin_layout Enumerate
Perform the operation.
\end_layout

\begin_layout Enumerate
Erase change log entries.
\end_layout

\begin_layout Enumerate
Unlock the file (or directory) on all of the lock servers.
\end_layout

\begin_layout Standard
The above algorithm is a simplified version intended for general users.
 Please refer to the source code for the full details.
\end_layout

\begin_layout Subsection*
Self-Heal
\end_layout

\begin_layout Standard
replicate automatically tries to fix any inconsistencies it detects among different
 copies of a file.
 It uses information in the change log to determine which copy is the 
\begin_inset Quotes eld
\end_inset

correct
\begin_inset Quotes erd
\end_inset

 version.
\end_layout

\begin_layout Standard
Self-heal is triggered when a file or directory is first 
\begin_inset Quotes eld
\end_inset

accessed
\begin_inset Quotes erd
\end_inset

, that is, the first time any operation is attempted on it.
 The self-heal algorithm does the following things:
\end_layout

\begin_layout Standard
If the entry being accessed is a directory:
\end_layout

\begin_layout Itemize
The contents of the 
\begin_inset Quotes eld
\end_inset

correct
\begin_inset Quotes erd
\end_inset

 version is replicated on all subvolumes, by deleting entries and creating
 entries as necessary.
\end_layout

\begin_layout Standard
If the entry being accessed is a file:
\end_layout

\begin_layout Itemize
If the file does not exist on some subvolumes, it is created.
\end_layout

\begin_layout Itemize
If there is a mismatch in the size of the file, or ownership, or permission,
 it is fixed.
\end_layout

\begin_layout Itemize
If the change log indicates that some copies need updating, they are updated.
\end_layout

\begin_layout Subsection*
Split-brain
\end_layout

\begin_layout Standard
It may happen that one replicate client can access only some of the servers in
 a cluster and another replicate client can access the remaining servers.
 Or it may happen that in a cluster of two servers, one server goes down
 and comes back up, but the other goes down immediately.
 Both these scenarios result in a 
\begin_inset Quotes eld
\end_inset

split-brain
\begin_inset Quotes erd
\end_inset

.
\end_layout

\begin_layout Standard
In a split-brain situation, there will be two or more copies of a file,
 all of which are 
\begin_inset Quotes eld
\end_inset

correct
\begin_inset Quotes erd
\end_inset

 in some sense.
 replicate without manual intervention has no way of knowing what to do, since
 it cannot consider any single copy as definitive, nor does it know of any
 meaningful way to merge the copies.
\end_layout

\begin_layout Standard
If replicate detects that a split-brain has happened on a file, it disallows opening
 of that file.
 You will have to manually resolve the conflict by deleting all but one
 copy of the file.
 Alternatively you can set an automatic split-brain resolution policy by
 using the `favorite-child' option (see below).
\end_layout

\begin_layout Section*
Translator Options
\end_layout

\begin_layout Standard
replicate accepts the following options:
\end_layout

\begin_layout Subsection*
read-subvolume (default: none)
\end_layout

\begin_layout Standard
The value of this option must be the name of a subvolume.
 If given, all read operations are sent to only the specified subvolume,
 instead of being balanced across all subvolumes.
\end_layout

\begin_layout Subsection*
favorite-child (default: none)
\end_layout

\begin_layout Standard
The value of this option must be the name of a subvolume.
 If given, the specified subvolume will be preferentially used in resolving
 conflicts (
\begin_inset Quotes eld
\end_inset

split-brain
\begin_inset Quotes erd
\end_inset

).
 This means if a discrepancy is noticed in the attributes or content of
 a file, the copy on the `favorite-child' will be considered the definitive
 version and its contents will 
\emph on
overwrite 
\emph default
the contents of all other copies.
 Use this option with caution! It is possible to 
\emph on
lose data
\emph default
 with this option.
 If you are in doubt, do not specify this option.
\end_layout

\begin_layout Subsection*
Self-heal options
\end_layout

\begin_layout Standard
Setting any of these options to 
\begin_inset Quotes eld
\end_inset

off
\begin_inset Quotes erd
\end_inset

 prevents that kind of self-heal from being done on a file or directory.
 For example, if metadata self-heal is turned off, permissions and ownership
 are no longer fixed automatically.
\end_layout

\begin_layout Subsubsection*
data-self-heal (default: on)
\end_layout

\begin_layout Standard
Enable/disable self-healing of file contents.
\end_layout

\begin_layout Subsubsection*
metadata-self-heal (default: off)
\end_layout

\begin_layout Standard
Enable/disable self-healing of metadata (permissions, ownership, modification
 times).
\end_layout

\begin_layout Subsubsection*
entry-self-heal (default: on)
\end_layout

\begin_layout Standard
Enable/disable self-healing of directory entries.
\end_layout

\begin_layout Subsection*
Change Log options
\end_layout

\begin_layout Standard
If any of these options is turned off, it disables writing of change log
 entries for that class of file operations.
 That is, steps 2 and 4 of the write algorithm (see above) are not done.
 Note that if the change log is not written, the self-heal algorithm cannot
 determine the 
\begin_inset Quotes eld
\end_inset

correct
\begin_inset Quotes erd
\end_inset

 version of a file and hence self-heal will only be able to fix 
\begin_inset Quotes eld
\end_inset

obviously
\begin_inset Quotes erd
\end_inset

 wrong things (such as a file existing on only one node).
\end_layout

\begin_layout Subsubsection*
data-change-log (default: on)
\end_layout

\begin_layout Standard
Enable/disable writing of change log for data operations.
\end_layout

\begin_layout Subsubsection*
metadata-change-log (default: on)
\end_layout

\begin_layout Standard
Enable/disable writing of change log for metadata operations.
\end_layout

\begin_layout Subsubsection*
entry-change-log (default: on)
\end_layout

\begin_layout Standard
Enable/disable writing of change log for entry operations.
\end_layout

\begin_layout Subsection*
Locking options
\end_layout

\begin_layout Standard
These options let you specify the number of lock servers to use for each
 class of file operations.
 The default values are satisfactory in most cases.
 If you are extra paranoid, you may want to increase the values.
 However, be very cautious if you set the data- or entry- lock server counts
 to zero, since this can result in 
\emph on
lost data.

\emph default
 For example, if you set the data-lock-server-count to zero, and two application
s write to the same region of a file, there is a possibility that none of
 your servers will have all the data.
 In other words, the copies will be 
\emph on
inconsistent
\emph default
, and 
\emph on
incomplete
\emph default
.
 Do not set data- and entry- lock server counts to zero unless you absolutely
 know what you are doing and agree to not hold GlusterFS responsible for
 any lost data.
\end_layout

\begin_layout Subsubsection*
data-lock-server-count (default: 1)
\end_layout

\begin_layout Standard
Number of lock servers to use for data operations.
\end_layout

\begin_layout Subsubsection*
metadata-lock-server-count (default: 0)
\end_layout

\begin_layout Standard
Number of lock servers to use for metadata operations.
\end_layout

\begin_layout Subsubsection*
entry-lock-server-count (default: 1)
\end_layout

\begin_layout Standard
Number of lock servers to use for entry operations.
\end_layout

\begin_layout Section*
Known Issues
\end_layout

\begin_layout Subsection*
Self-heal of file with more than one link (hard links):
\end_layout

\begin_layout Standard
Consider two servers, A and B.
 Assume A is down, and the user creates a file `new' as a hard link to a
 file `old'.
 When A comes back up, replicate will see that the file `new' does not exist on
 A, and self-heal will create the file and copy the contents from B.
 However, now on server A the file `new' is not a link to the file `old'
 but an entirely different file.
\end_layout

\begin_layout Standard
We know of no easy way to fix this problem, but we will try to fix it in
 forthcoming releases.
\end_layout

\begin_layout Subsection*
File re-opening after a server comes back up:
\end_layout

\begin_layout Standard
If a server A goes down and comes back up, any files which were opened while
 A was down and are still open will not have their writes replicated on
 A.
 In other words, data replication only happens on those servers which were
 alive when the file was opened.
\end_layout

\begin_layout Standard
This is a rather tricky issue but we hope to fix it very soon.
\end_layout

\begin_layout Section*
Frequently Asked Questions
\end_layout

\begin_layout Subsection*
1.
 How can I force self-heal to happen?
\end_layout

\begin_layout Standard
You can force self-heal to happen on your cluster by running a script or
 a command that accesses every file.
 A simple way to do it would be:
\end_layout

\begin_layout Standard
\begin_inset ERT
status open

\begin_layout Standard

\end_layout

\begin_layout Standard


\backslash
begin{verbatim}
\end_layout

\begin_layout Standard

$ ls -lR
\end_layout

\begin_layout Standard


\backslash
end{verbatim}
\end_layout

\begin_layout Standard

\end_layout

\end_inset


\end_layout

\begin_layout Standard
Run the command in all directories which you want to forcibly self-heal.
\end_layout

\begin_layout Subsection*
2.
 Which backend filesystem should I use for replicate?
\end_layout

\begin_layout Standard
You can use any backend filesystem that supports extended attributes.
 We know of users successfully using XFR, ReiserFS, and Ext3.
\end_layout

\begin_layout Subsection*
3.
 What can I do to improve replicate performance?
\end_layout

\begin_layout Standard
Try loading performance translators such as io-threads, write-behind, io-cache,
 and read-ahead depending on your workload.
 If you are willing to sacrifice correctness in corner cases, you can experiment
 with the lock-server-count and the change-log options (see above).
 As warned earlier, be very careful!
\end_layout

\begin_layout Subsection*
4.
 How can I selectively replicate files?
\end_layout

\begin_layout Standard
There is no support for selective replication in replicate itself.
 You can achieve selective replication by loading the unify translator over
 replicate, and using the switch scheduler.
 Configure unify with two subvolumes, one of them being replicate.
 Using the switch scheduler, schedule all files for which you need replication
 to the replicate subvolume.
 Consult unify and switch documentation for more details.
\end_layout

\begin_layout Section*
Contact
\end_layout

\begin_layout Standard
If you need more assistance on replicate, contact us on the mailing list <gluster-user
s@gluster.org> (visit gluster.org for details on how to subscribe).
\end_layout

\begin_layout Standard
Send you comments and suggestions about this document to <vikas@gluster.com>.
\end_layout

\end_body
\end_document