On 05/22/2012 02:11 AM, Anand Avati wrote:
On Tue, May 8, 2012 at 2:34 AM, Xavier Hernandez
<[email protected] <mailto:[email protected]>> wrote:
Hello developers,
I would like to expose some ideas we are working on to create a
new kind of translator that should be able to unify and simplify
to some extent the healing procedures of complex translators.
Currently, the only translator with complex healing capabilities
that we are aware of is AFR. We are developing another translator
that will also need healing capabilities, so we thought that it
would be interesting to create a new translator able to handle the
common part of the healing process and hence to simplify and avoid
duplicated code in other translators.
The basic idea of the new translator is to handle healing tasks
nearer the storage translator on the server nodes instead to
control everything from a translator on the client nodes. Of
course the heal translator is not able to handle healing entirely
by itself, it needs a client translator which will coordinate all
tasks. The heal translator is intended to be used by translators
that work with multiple subvolumes.
I will try to explain how it works without entering into too much
details.
There is an important requisite for all client translators that
use healing: they must have exactly the same list of subvolumes
and in the same order. Currently, I think this is not a problem.
The heal translator treats each file as an independent entity, and
each one can be in 3 modes:
1. Normal mode
This is the normal mode for a copy or fragment of a file when
it is synchronized and consistent with the same file on other
nodes (for example with other replicas. It is the client
translator who decides if it is synchronized or not).
2. Healing mode
This is the mode used when a client detects an inconsistency
in the copy or fragment of the file stored on this node and
initiates the healing procedures.
3. Provider mode (I don't like very much this name, though)
This is the mode used by client translators when an
inconsistency is detected in this file, but the copy or
fragment stored in this node is considered good and it will be
used as a source to repair the contents of this file on other
nodes.
Initially, when a file is created, it is set in normal mode.
Client translators that make changes must guarantee that they send
the modification requests in the same order to all the servers.
This should be done using inodelk/entrylk.
When a change is sent to a server, the client must include a
bitmap mask of the clients to which the request is being sent.
Normally this is a bitmap containing all the clients, however,
when a server fails for some reason some bits will be cleared. The
heal translator uses this bitmap to early detect failures on other
nodes from the point of view of each client. When this condition
is detected, the request is aborted with an error and the client
is notified with the remaining list of valid nodes. If the client
considers the request can be successfully server with the
remaining list of nodes, it can resend the request with the
updated bitmap.
The heal translator also updates two file attributes for each
change request to mantain the "version" of the data and metadata
contents of the file. A similar task is currently made by AFR
using xattrop. This would not be needed anymore, speeding write
requests.
The version of data and metadata is returned to the client for
each read request, allowing it to detect inconsistent data.
When a client detects an inconsistency, it initiates healing.
First of all, it must lock the entry and inode (when necessary).
Then, from the data collected from each node, it must decide which
nodes have good data and which ones have bad data and hence need
to be healed. There are two possible cases:
1. File is not a regular file
In this case the reconstruction is very fast and requires few
requests, so it is done while the file is locked. In this
case, the heal translator does nothing relevant.
2. File is a regular file
For regular files, the first step is to synchronize the
metadata to the bad nodes, including the version information.
Once this is done, the file is set in healing mode on bad
nodes, and provider mode on good nodes. Then the entry and
inode are unlocked.
When a file is in provider mode, it works as in normal mode, but
refuses to start another healing. Only one client can be healing a
file.
When a file is in healing mode, each normal write request from any
client are handled as if the file were in normal mode, updating
the version information and detecting possible inconsistencies
with the bitmap. Additionally, the healing translator marks the
written region of the file as "good".
Each write request from the healing client intended to repair the
file must be marked with a special flag. In this case, the area
that wants to be written is filtered by the list of "good" ranges
(if there are any intersection with a good range, it is removed
from the request). The resulting set of ranges are propagated to
the lower translator and added to the list of "good" ranges but
the version information is not updated.
Read requests are only served if the range requested is entirely
contained into the "good" regions list.
There are some additional details, but I think this is enough to
have a general idea of its purpose and how it works.
The main advantages of this translator are:
1. Avoid duplicated code in client translators
2. Simplify and unify healing methods in client translators
3. xattrop is not needed anymore in client translators to keep
track of changes
4. Full file contents are repaired without locking the file
5. Better detection and prevention of some split brain situations
as soon as possible
I think it would be very useful. It seems to me that it works
correctly in all situations, however I don't have all the
experience that other developers have with the healing functions
of AFR, so I will be happy to answer any question or suggestion to
solve problems it may have or to improve it.
What do you think about it ?
The goals you state above are all valid. What would really help
(adoption) is if you can implement this as a modification of AFR by
utilizing all the work already done, and you get brownie points if it
is backward compatible with existing AFR. If you already have any code
in a publishable state, please share it with us (github link?).
Avati
I've tried to understand how AFR works and, in some way, some of the
ideas have been taken from it. However it is very complex and a lot of
changes have been carried out in the master branch over the latest
months. It's hard for me to follow them while actively working on my
translator. Nevertheless, the main reason to take a separate path was
that AFR is strongly bound to replication (at least from what I saw when
I analyzed it more deeply. Maybe things have changed now, but haven't
had time to review them).
The requirements for my translator didn't fit very well with AFR, and
the needed effort to understand and modify it to adapt it was too high.
It also seems that there isn't any detailed developer info about
internals of AFR that could have helped to be more confident to modify
it (at least I haven't found it).
I'm currenty working on it, but it's not ready yet. As soon as it is in
a minimally stable state we will publish it, probably on github. I'll
write the url to this list.
Thank you
_______________________________________________
Gluster-devel mailing list
[email protected]
https://lists.nongnu.org/mailman/listinfo/gluster-devel
