zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: Efficient backup and a reasonable restore of an ensemble
Date Tue, 09 Jul 2013 05:32:34 GMT

It isn't that bad.  The deal is that a snapshot takes time to write to disk.  During this
time updates are still allowed to the contents of memory.  All such updates are logged however
so if you have the transaction log from the moment before the snap starts until some moment
after the snap completes you can load the snapshot and then replay the log to get a moment
in time snapshot as of the time if the final transaction that you have applied.  

This works because all if the logged transactions are idem potent. If they are applied to
part of the snapshot that already recorded their effect, there is no problem. 

If you want you can even do the replay in a side process after the snapshot is complete so
that you don't have to carry around the transaction log.  

Sent from my iPhone

On Jul 8, 2013, at 21:42, Sergey Maslyakov <evolvah@gmail.com> wrote:

> Kishore,
> This sounds like a very elaborate tool. I was trying to find a simplistic
> approach but what Thawan said about "fuzzy snapshots" makes me a little
> afraid that there is no simple solution.
> On Mon, Jul 8, 2013 at 11:05 PM, kishore g <g.kishore@gmail.com> wrote:
>> Agree, we already have such a tool. In fact we use it to reconstruct the
>> sequence of events that led to a failure and actually restore the system to
>> a previous stable point and replay the events. Unfortunately this is tied
>> closely with Helix but it should be easy to make this a generic tool.
>> Sergey is this something that will be useful in your case.
>> Thanks,
>> Kishore G
>> On Mon, Jul 8, 2013 at 8:09 PM, Thawan Kooburat <thawan@fb.com> wrote:
>>> On restore part, I think having a separate utility to manipulate the
>>> data/snap dir (by truncating the log/removing snapshot to a given zxid)
>>> would be easier than modifying the server.
>>> --
>>> Thawan Kooburat
>>> On 7/8/13 6:34 PM, "kishore g" <g.kishore@gmail.com> wrote:
>>>> I think what we are looking at is a  point in time restore
>> functionality.
>>>> How about adding a feature that says go back to a specific
>> zxid/timestamp.
>>>> This way before doing any change to zookeeper simply note down the
>>>> timestamp/zxid on leader. If things go wrong after making changes, bring
>>>> down zookeepers and provide additional parameter of a zxid/timestamp
>> while
>>>> restarting. The server can go the exact point and make it current. The
>>>> followers can be started blank.
>>>> On Mon, Jul 8, 2013 at 5:53 PM, Thawan Kooburat <thawan@fb.com> wrote:
>>>>> Just saw that  this is the corresponding use case to the question
>> posted
>>>>> in dev list.
>>>>> In order to restore the data to a given point in time correctly, you
>>>>> need
>>>>> both snapshot and txnlog. This is because zookeeper snapshot is fuzzy
>>>>> and
>>>>> snapshot alone may not represent a valid state of the server if there
>>>>> are
>>>>> in-flight requests.
>>>>> The 4wl command should cause the server to roll the log and take a
>>>>> snapshot similar to periodic snapshotting operation. Your backup
>> script
>>>>> need grap the snapshot and corresponding txnlog file from the data
>> dir.
>>>>> To restore, just shutdown all hosts, clear the data dir, copy over the
>>>>> snapshot and txnlog, and restart them.
>>>>> --
>>>>> Thawan Kooburat
>>>>> On 7/8/13 3:28 PM, "Sergey Maslyakov" <evolvah@gmail.com> wrote:
>>>>>> Thank you for your response, Flavio. I apologize, I did not provide
>>>>>> clear
>>>>>> explanation of the use case.
>>>>>> This backup/restore is not intended to be tied to any write event,
>>>>>> instead,
>>>>>> it is expected to run as a periodic (daily?) cron job on one of the
>>>>>> servers, which is not guaranteed to be the leader of the ensemble.
>>>>> There
>>>>>> is
>>>>>> no expectation that all recent changes are committed and persisted
>>>>>> disk.
>>>>>> The system can sustain the loss of several hours worth of recent
>>>>> changes
>>>>>> in
>>>>>> the event of restore.
>>>>>> As for finding the leader dynamically and performing backup on it,
>> this
>>>>>> approach could be more difficult as the leader can change time to
>> time
>>>>> and
>>>>>> I still need to fetch the file to store it in my designated backup
>>>>>> location. Taking backup on one server and picking it up from a local
>>>>> file
>>>>>> system looks less error-prone. Even if I went the fancy route and
>>>>>> Zookeeper send me the serialized DataTree in response to the 4wl,
>> this
>>>>>> approach would involve a lot of moving parts.
>>>>>> I have already made a PoC for a new 4wl that invokes takeSnapshot()
>> and
>>>>>> returns an absolute path to the snapshot it drops on disk. I have
>>>>> already
>>>>>> protected takeSnapshot() from concurrent invocation, which is likely
>> to
>>>>>> corrupt the snapshot file on disk. This approach works but I'm
>>>>> thinking to
>>>>>> take it one step further by providing the desired path name as an
>>>>> argument
>>>>>> to my new 4lw and to have Zookeeper server drop the snapshot into
>>>>>> specified file and report success/failure back. This way I can avoid
>>>>>> cluttering the data directory and interfering with what Zookeeper
>> finds
>>>>>> when it scans the data directory.
>>>>>> Approach with having an additional server that would take the
>>>>> leadership
>>>>>> and populate the ensemble is just a theory. I don't see a clean way
>> of
>>>>>> making a quorum member the leader of the quorum. Am I overlooking
>>>>>> something
>>>>>> simple?
>>>>>> In backup and restore of an ensemble the biggest unknown for me
>> remains
>>>>>> populating the ensemble with desired data. I can think of two ways:
>>>>>> 1. Clear out all servers by stopping them, purge version-2
>> directories,
>>>>>> restore a snapshot file on one server that will be brought first,
>>>>> then
>>>>>> bring up the rest of the ensemble. This way I somewhat force the
>> first
>>>>>> server to be the leader because it has data and it will be the only
>>>>> member
>>>>>> of a quorum with data, provided to the way I start the ensemble.
>>>>>> looks
>>>>>> like a hack, though.
>>>>>> 2. Clear out the ensemble and reload it with a dedicated client using
>>>>> the
>>>>>> provided Zookeeper API.
>>>>>> With the approach of backing up an actual snapshot file, option #1
>>>>> appears
>>>>>> to be more practical.
>>>>>> I wish I could start the ensemble with a designate leader that would
>>>>>> bootstrap the ensemble with data and then the ensemble would go into
>>>>> its
>>>>>> normal business...
>>>>>> On Mon, Jul 8, 2013 at 4:30 PM, Flavio Junqueira
>>>>>> <fpjunqueira@yahoo.com>wrote:
>>>>>>> One bit that is still a bit confusing to me in your use case
is if
>>>>> you
>>>>>>> need to take a snapshot right after some event in your application.
>>>>>>> Even if
>>>>>>> you're able to tell ZooKeeper to take a snapshot, there is no
>>>>> guarantee
>>>>>>> that it will happen at the exact point you want it if update
>>>>> operations
>>>>>>> keep coming.
>>>>>>> If you use your four-letter word approach, then would you search
>> for
>>>>> the
>>>>>>> leader or would you simply take a snapshot at any server? If
it has
>>>>> to
>>>>>>> go
>>>>>>> through the leader so that you make sure to have the most recent
>>>>>>> committed
>>>>>>> state, then it might not be a bad idea to have an api call that
>> tells
>>>>>>> the
>>>>>>> leader to take a snapshot at some directory of your choice.
>> Informing
>>>>>>> you
>>>>>>> the name of the snapshot file so that you can copy sounds like
>>>>>>> option,
>>>>>>> but perhaps it is not as convenient.
>>>>>>> The approach of adding another server is not very clear. How
do you
>>>>>>> force
>>>>>>> it to be the leader? Keep in mind that if it crashes, then it
>>>>> lose
>>>>>>> leadership.
>>>>>>> -Flavio
>>>>>>> On Jul 8, 2013, at 8:34 AM, Sergey Maslyakov <evolvah@gmail.com>
>>>>> wrote:
>>>>>>>> It looks like the "dev" mailing list is rather inactive.
Over the
>>>>> past
>>>>>>> few
>>>>>>>> days I only saw several automated emails from JIRA and this
>>>>> pretty
>>>>>>> much
>>>>>>>> it. Contrary to this, the "user" mailing list seems to be
>>>>> alive
>>>>>>> and
>>>>>>>> more populated.
>>>>>>>> With this in mind, please allow me to cross-post here the
>> message I
>>>>>>> sent
>>>>>>>> into the "dev" list a few days ago.
>>>>>>>> Regards,
>>>>>>>> /Sergey
>>>>>>>> === forwarded message begins here ===
>>>>>>>> Hi!
>>>>>>>> I'm facing the problem that has been raised by multiple people
>> but
>>>>>>> none
>>>>>>> of
>>>>>>>> the discussion threads seem to provide a good answer. I dug
>>>>>>> Zookeeper
>>>>>>>> source code trying to come up with some possible approaches
and I
>>>>>>> would
>>>>>>>> like to get your inputs on those.
>>>>>>>> Initial conditions:
>>>>>>>> * I have an ensemble of five Zookeeper servers running v3.4.5
>> code.
>>>>>>>> * The size of a committed snapshot file is in vicinity of
>>>>>>>> * There are about 80 clients connected to the ensemble.
>>>>>>>> * Clients a heavily read biased, i.e., they mostly read and
>> rarely
>>>>>>> write. I
>>>>>>>> would say less than 0.1% of queries modify the data.
>>>>>>>> Problem statement:
>>>>>>>> * Under certain conditions, I may need to revert the data
>> in
>>>>>>> the
>>>>>>>> ensemble to an earlier state. For example, one of the clients
>>>>> ruin
>>>>>>> the
>>>>>>>> application-level data integrity and I need to perform a
>>>>>>> recovery.
>>>>>>>> Things look nice and easy if I'm dealing with a single Zookeeper
>>>>>>> server.
>>>>>>> A
>>>>>>>> file-level copy of the data and dataLog directories should
>>>>> me to
>>>>>>>> recover later by stopping Zookeeper, swapping the corrupted
>>>>> and
>>>>>>>> dataLog directories with a backup, and firing Zookeeper back
>>>>>>>> Now, the ensemble deployment and the leader election algorithm
>>>>> the
>>>>>>>> quorum make things much more difficult. In order to restore
>> a
>>>>>>> single
>>>>>>>> file-level backup, I need to take the whole ensemble down,
>> out
>>>>>>> data
>>>>>>>> and dataLog directories on all servers, replace these directories
>>>>> with
>>>>>>>> backed up content on one of the servers, bring this server
>>>>> first,
>>>>>>> and
>>>>>>>> then bring up the rest of the ensemble. This [somewhat]
>> guarantees
>>>>>>> that
>>>>>>> the
>>>>>>>> populated Zookeeper server becomes a member of a majority
>>>>>>> populates
>>>>>>> the
>>>>>>>> ensemble. This approach works but it is very involving and,
>>>>>>>> error-prone due to a human error.
>>>>>>>> Based on a study of Zookeeper source code, I am considering
>>>>>>> following
>>>>>>>> alternatives. And I seek advice from Zookeeper development
>>>>> community
>>>>>>> as
>>>>>>> to
>>>>>>>> which approach looks more promising or if there is a better
>>>>>>>> Approach #1:
>>>>>>>> Develop a complementary pair of utilities for export and
>> of
>>>>> the
>>>>>>>> data. Both utilities will act as Zookeeper clients and use
>>>>>>> existing
>>>>>>>> API. The "export" utility will recursively retrieve data
>> store
>>>>> it
>>>>>>> in
>>>>>>> a
>>>>>>>> file. The "import" utility will first purge all data from
>>>>> ensemble
>>>>>>> and
>>>>>>>> then reload it from the file.
>>>>>>>> This approach seems to be the simplest and there are similar
>> tools
>>>>>>>> developed already. For example, the Guano Project:
>>>>>>>> https://github.com/d2fn/guano
>>>>>>>> I don't like two things about it:
>>>>>>>> * Poor performance even on a backup for the data store of
>> size.
>>>>>>>> * Possible data consistency issues due to concurrent access
>> the
>>>>>>> export
>>>>>>>> utility as well as other "normal" clients.
>>>>>>>> Approach #2:
>>>>>>>> Add another four-letter command that would force rolling
up the
>>>>>>>> transactions and creating a snapshot. The result of this
>>>>> would
>>>>>>> be a
>>>>>>>> new snapshot.XXXX file on disk and the name of the file could
>>>>>>> reported
>>>>>>>> back to the client as a response to the four-letter command.
>>>>>>> way, I
>>>>>>>> would know which snapshot file to grab for future possible
>> restore.
>>>>>>> But
>>>>>>>> restoring from a snapshot file is almost as involving as
>>>>>>> error-prone
>>>>>>>> sequence described in the "Initial conditions" above.
>>>>>>>> Approach #3:
>>>>>>>> Come up with a way to temporarily add a new Zookeeper server
>> into a
>>>>>>> live
>>>>>>>> ensemble, that would overtake (how?) the leader role and
push out
>>>>> the
>>>>>>>> snapshot that it has into all ensemble members upon restore.
>>>>>>> approach
>>>>>>>> could be difficult and error-prone to implement because it
>>>>>>> require
>>>>>>>> hacking the existing election algorithm to designate a leader.
>>>>>>>> So, which of the approaches do you think works best for an
>> ensemble
>>>>>>> and
>>>>>>> for
>>>>>>>> the database size of about 1GB?
>>>>>>>> Any advice will be highly appreciated!
>>>>>>>> /Sergey

View raw message