Raft Refloated
88 views
Skip to first unread message
Kijana Woodard
Feb 10, 2015, 1:00:37 PM2/10/15
to raft...@googlegroups.com
Heidi Howard
Feb 10, 2015, 1:06:18 PM2/10/15
to Kijana Woodard, raft...@googlegroups.com
I'd be keen to hear them :)
On 10 February 2015 at 18:00, Kijana Woodard <kijana....@gmail.com> wrote:
--
You received this message because you are subscribed to the Google Groups "raft-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to raft-dev+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
Regards
Heidi
Heidi
Kijana Woodard
Feb 11, 2015, 11:34:45 AM2/11/15
to Heidi Howard, raft...@googlegroups.com
I'll start with a light comment.
"It would have helped us keep track if an explicit “change log” had been available although the original research was still ongoing at the time"
Would it be as simple as writing the drafts in a text format that git[hub] can do diffs upon?
Then one could use "compare view" to get a combined diff of changes since a previous version.
Skype Xu
Feb 11, 2015, 11:03:20 PM2/11/15
to raft...@googlegroups.com, hh...@cam.ac.uk
```
Raft addresses
this problem by restricting the leader to only commit entries
from any previous term if the same leader has already successfully
replicated an entry from the current term. This is later referred to
as the “extra condition on commitment”.```
This sentence is quite confusing even in original raft paper.
I can not understand why a new leader can replicate a log in current term while previous term log is not replicated successfully,
won't prevLogIndex and prevLogTerm prevent it from replicating current term log successfully ?
Figure 6 is confusing, how a new leader avoid this problem ?
Regards
Kijana Woodard
Feb 11, 2015, 11:16:02 PM2/11/15
to Skype Xu, raft...@googlegroups.com, hh...@cam.ac.uk
Commit not replicate.
The entries from previous terms _and_ at least one entry from the current leader's term have to be replicated to a majority of the cluster in order to commit.
The entries from previous terms _and_ at least one entry from the current leader's term have to be replicated to a majority of the cluster in order to commit.
From: Skype Xu
Sent: 2/11/2015 10:03 PM
To: raft...@googlegroups.com
Cc: hh...@cam.ac.uk
Subject: Re: Raft Refloated
Diego Ongaro
Feb 13, 2015, 2:10:31 PM2/13/15
to Kijana Woodard, Heidi Howard, raft...@googlegroups.com
In retrospect, we should have just kept a changelog. I guess it
surprised us that the pre-publication version of the paper was so
popular, and also that it took so long to publish.
I think it would have taken a proper changelog, since the textual
diffs would have been cluttered with tons of noise from the many times
we rewrote and reorganized the text for clarity. Also, there's SVGs
and a PPTX in there (for Figure 2) that aren't easy to diff. This
might give you some idea of the churn on the *.tex files:
$ wc -l <*.tex
2557
$ git log --shortstat --pretty=format: *.tex | awk '{ins += $4; del +=
$6} END {print ins " insertions, " del " deletions"}'
11420 insertions, 8743 deletions
-Diego
surprised us that the pre-publication version of the paper was so
popular, and also that it took so long to publish.
I think it would have taken a proper changelog, since the textual
diffs would have been cluttered with tons of noise from the many times
we rewrote and reorganized the text for clarity. Also, there's SVGs
and a PPTX in there (for Figure 2) that aren't easy to diff. This
might give you some idea of the churn on the *.tex files:
$ wc -l <*.tex
2557
$ git log --shortstat --pretty=format: *.tex | awk '{ins += $4; del +=
$6} END {print ins " insertions, " del " deletions"}'
11420 insertions, 8743 deletions
-Diego
Reply all
Reply to author
Forward
0 new messages