You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@zookeeper.apache.org by jiangwen w <wj...@gmail.com> on 2011/04/06 06:31:05 UTC
discarded messages seems will exists in some follower, when the
leader failed.
discarded messages seems will exists in some follower, when the leader
failed. considering the following corner case:
the latest committed proposals in the New Leader. the following is the
latest zxid list with format: *<epoch>,<lower bits>*
1,7
1,8
2,0
2,1
the last zxid in a follower
1,7
1,8
1,9
when the follower sync with the new leader, the transaction *1,9 *should be
truncated, but I couldn't found the code that can handle this case.
In this case the following code in LeanerHandler seems will be executed.
if (proposals.size() != 0) {
if ((leader.zk.getZKDatabase().getmaxCommittedLog() >=
peerLastZxid)
&&
(leader.zk.getZKDatabase().getminCommittedLog() <= peerLastZxid)) {
packetToSend = Leader.DIFF;
Sincerely
Re: discarded messages seems will exists in some follower, when the
leader failed.
Posted by jiangwen w <wj...@gmail.com>.
for case 1:
how can the inequity minCommitedLog of Y > 1,9 hold? in your case
minCommitedLog
of Y is 0. by the way minCommitedLog of Y will not be, because the new
leader will load the snapshot and replay the transaction logs.
for case 2:
follower append proposals to transaction log before proposals are
committed.
so if follower logging 1,9. the leader may have not committed 1,9 to the
quorum
Sincerely
On Thu, Apr 7, 2011 at 10:53 PM, Vishal Kher <vi...@gmail.com> wrote:
> Looking at the code, I don't think this can happen.
>
>
> > > the latest committed proposals in the New Leader. the following is the
> > > latest zxid list with format: *<epoch>,<lower bits>*
> > > 1,7
> > > 1,8
> > > 2,0
> > > 2,1
> > >
> > > the last zxid in a follower 'Z'
>
> > > 1,7
> > > 1,8
> > > 1,9
>
> when can follower Z have 1,9 in its log?
>
> case 1: Z was the leader and proposed 1,9. But died before sending it to
> other peers.
> In this case a new leader (say Y) will have to be elected. Prior to
> election,Y will call follower.shutdown() which cleanup everything and reset
> minCommittedLog to 0 (ZKDatabase.clear()).
> Later, when Z attempts to join Y, Y will send TRUNC to Z, because
> minCommitedLog of Y > 1,9.
>
> case 2: Z was follower and died after logging 1,9.
> In this case the leader will have commited 1,9 to the quorum.
>
>
>
> On Thu, Apr 7, 2011 at 4:46 AM, jiangwen w <wj...@gmail.com> wrote:
>
>> Hi Mahadev,
>>
>> thanks for your answer!
>> the leader send TRUNC when the last ZXID in the follower is larger than
>> the
>> max committed ZXID in the leader.
>> but in this case the last ZXID in the follower is 1,9 the max committed
>> ZXID
>> in the leader is 2,1. So the leader will not send the TRUNC.
>> the following is the code in LeanerHandler:
>>
>> if (logTxns && (peerLastZxid >
>> leader.zk.getZKDatabase().getmaxCommittedLog())) {
>> // this is the only case that we are sure that
>> // we can ask the peer to truncate the log
>> packetToSend = Leader.TRUNC;
>> zxidToSend =
>> leader.zk.getZKDatabase().getmaxCommittedLog();
>> updates = zxidToSend;
>> }
>>
>> Sincerely
>>
>>
>>
>> On Thu, Apr 7, 2011 at 11:21 AM, Mahadev Konar <ma...@apache.org>
>> wrote:
>>
>> > Hi Jiangwen,
>> > Take a look at :
>> >
>> > syncWithLeader(newLeaderZxid);
>> >
>> > This method calls truncate to take care of the issue you mention.
>> >
>> > thanks
>> > mahadev
>> >
>> > On Tue, Apr 5, 2011 at 9:31 PM, jiangwen w <wj...@gmail.com> wrote:
>> > > discarded messages seems will exists in some follower, when the leader
>> > > failed. considering the following corner case:
>> > >
>> > > the latest committed proposals in the New Leader. the following is the
>> > > latest zxid list with format: *<epoch>,<lower bits>*
>> > > 1,7
>> > > 1,8
>> > > 2,0
>> > > 2,1
>> > >
>> > > the last zxid in a follower
>> > > 1,7
>> > > 1,8
>> > > 1,9
>> > >
>> > > when the follower sync with the new leader, the transaction *1,9
>> *should
>> > be
>> > > truncated, but I couldn't found the code that can handle this case.
>> > > In this case the following code in LeanerHandler seems will be
>> executed.
>> > >
>> > > if (proposals.size() != 0) {
>> > > if ((leader.zk.getZKDatabase().getmaxCommittedLog()
>> >=
>> > > peerLastZxid)
>> > > &&
>> > > (leader.zk.getZKDatabase().getminCommittedLog() <= peerLastZxid)) {
>> > > packetToSend = Leader.DIFF;
>> > >
>> > > Sincerely
>> > >
>> >
>> >
>> >
>> > --
>> > thanks
>> > mahadev
>> > @mahadevkonar
>> >
>>
>
>
Re: discarded messages seems will exists in some follower, when the
leader failed.
Posted by Vishal Kher <vi...@gmail.com>.
Looking at the code, I don't think this can happen.
> > the latest committed proposals in the New Leader. the following is the
> > latest zxid list with format: *<epoch>,<lower bits>*
> > 1,7
> > 1,8
> > 2,0
> > 2,1
> >
> > the last zxid in a follower 'Z'
> > 1,7
> > 1,8
> > 1,9
when can follower Z have 1,9 in its log?
case 1: Z was the leader and proposed 1,9. But died before sending it to
other peers.
In this case a new leader (say Y) will have to be elected. Prior to
election,Y will call follower.shutdown() which cleanup everything and reset
minCommittedLog to 0 (ZKDatabase.clear()).
Later, when Z attempts to join Y, Y will send TRUNC to Z, because
minCommitedLog of Y > 1,9.
case 2: Z was follower and died after logging 1,9.
In this case the leader will have commited 1,9 to the quorum.
On Thu, Apr 7, 2011 at 4:46 AM, jiangwen w <wj...@gmail.com> wrote:
> Hi Mahadev,
>
> thanks for your answer!
> the leader send TRUNC when the last ZXID in the follower is larger than the
> max committed ZXID in the leader.
> but in this case the last ZXID in the follower is 1,9 the max committed
> ZXID
> in the leader is 2,1. So the leader will not send the TRUNC.
> the following is the code in LeanerHandler:
>
> if (logTxns && (peerLastZxid >
> leader.zk.getZKDatabase().getmaxCommittedLog())) {
> // this is the only case that we are sure that
> // we can ask the peer to truncate the log
> packetToSend = Leader.TRUNC;
> zxidToSend = leader.zk.getZKDatabase().getmaxCommittedLog();
> updates = zxidToSend;
> }
>
> Sincerely
>
>
>
> On Thu, Apr 7, 2011 at 11:21 AM, Mahadev Konar <ma...@apache.org> wrote:
>
> > Hi Jiangwen,
> > Take a look at :
> >
> > syncWithLeader(newLeaderZxid);
> >
> > This method calls truncate to take care of the issue you mention.
> >
> > thanks
> > mahadev
> >
> > On Tue, Apr 5, 2011 at 9:31 PM, jiangwen w <wj...@gmail.com> wrote:
> > > discarded messages seems will exists in some follower, when the leader
> > > failed. considering the following corner case:
> > >
> > > the latest committed proposals in the New Leader. the following is the
> > > latest zxid list with format: *<epoch>,<lower bits>*
> > > 1,7
> > > 1,8
> > > 2,0
> > > 2,1
> > >
> > > the last zxid in a follower
> > > 1,7
> > > 1,8
> > > 1,9
> > >
> > > when the follower sync with the new leader, the transaction *1,9
> *should
> > be
> > > truncated, but I couldn't found the code that can handle this case.
> > > In this case the following code in LeanerHandler seems will be
> executed.
> > >
> > > if (proposals.size() != 0) {
> > > if ((leader.zk.getZKDatabase().getmaxCommittedLog()
> >=
> > > peerLastZxid)
> > > &&
> > > (leader.zk.getZKDatabase().getminCommittedLog() <= peerLastZxid)) {
> > > packetToSend = Leader.DIFF;
> > >
> > > Sincerely
> > >
> >
> >
> >
> > --
> > thanks
> > mahadev
> > @mahadevkonar
> >
>
Re: discarded messages seems will exists in some follower, when the
leader failed.
Posted by jiangwen w <wj...@gmail.com>.
Hi Mahadev,
thanks for your answer!
the leader send TRUNC when the last ZXID in the follower is larger than the
max committed ZXID in the leader.
but in this case the last ZXID in the follower is 1,9 the max committed ZXID
in the leader is 2,1. So the leader will not send the TRUNC.
the following is the code in LeanerHandler:
if (logTxns && (peerLastZxid >
leader.zk.getZKDatabase().getmaxCommittedLog())) {
// this is the only case that we are sure that
// we can ask the peer to truncate the log
packetToSend = Leader.TRUNC;
zxidToSend = leader.zk.getZKDatabase().getmaxCommittedLog();
updates = zxidToSend;
}
Sincerely
On Thu, Apr 7, 2011 at 11:21 AM, Mahadev Konar <ma...@apache.org> wrote:
> Hi Jiangwen,
> Take a look at :
>
> syncWithLeader(newLeaderZxid);
>
> This method calls truncate to take care of the issue you mention.
>
> thanks
> mahadev
>
> On Tue, Apr 5, 2011 at 9:31 PM, jiangwen w <wj...@gmail.com> wrote:
> > discarded messages seems will exists in some follower, when the leader
> > failed. considering the following corner case:
> >
> > the latest committed proposals in the New Leader. the following is the
> > latest zxid list with format: *<epoch>,<lower bits>*
> > 1,7
> > 1,8
> > 2,0
> > 2,1
> >
> > the last zxid in a follower
> > 1,7
> > 1,8
> > 1,9
> >
> > when the follower sync with the new leader, the transaction *1,9 *should
> be
> > truncated, but I couldn't found the code that can handle this case.
> > In this case the following code in LeanerHandler seems will be executed.
> >
> > if (proposals.size() != 0) {
> > if ((leader.zk.getZKDatabase().getmaxCommittedLog() >=
> > peerLastZxid)
> > &&
> > (leader.zk.getZKDatabase().getminCommittedLog() <= peerLastZxid)) {
> > packetToSend = Leader.DIFF;
> >
> > Sincerely
> >
>
>
>
> --
> thanks
> mahadev
> @mahadevkonar
>
Re: discarded messages seems will exists in some follower, when the
leader failed.
Posted by Mahadev Konar <ma...@apache.org>.
Hi Jiangwen,
Take a look at :
syncWithLeader(newLeaderZxid);
This method calls truncate to take care of the issue you mention.
thanks
mahadev
On Tue, Apr 5, 2011 at 9:31 PM, jiangwen w <wj...@gmail.com> wrote:
> discarded messages seems will exists in some follower, when the leader
> failed. considering the following corner case:
>
> the latest committed proposals in the New Leader. the following is the
> latest zxid list with format: *<epoch>,<lower bits>*
> 1,7
> 1,8
> 2,0
> 2,1
>
> the last zxid in a follower
> 1,7
> 1,8
> 1,9
>
> when the follower sync with the new leader, the transaction *1,9 *should be
> truncated, but I couldn't found the code that can handle this case.
> In this case the following code in LeanerHandler seems will be executed.
>
> if (proposals.size() != 0) {
> if ((leader.zk.getZKDatabase().getmaxCommittedLog() >=
> peerLastZxid)
> &&
> (leader.zk.getZKDatabase().getminCommittedLog() <= peerLastZxid)) {
> packetToSend = Leader.DIFF;
>
> Sincerely
>
--
thanks
mahadev
@mahadevkonar