Sorry I couldn't make it up for the powwow.
Thanks Devaraj for the excellent notes.
On this part:
-- Branching on a big feature was suggested -
--- Issues: committer needs to be there
Sergey: If the rate of change is high on common code (to
the branches)
then merging will be tough
--- Jon: Refactor should be done in the main branch (since it doesn't
add any new funtionality)
I like how snapshots was branched and worked on collaboratively up in
GitHub.
It's pretty painless to rebase with git but painful to keep a branch up to
date with subversion then merge back (at least for me).
I have something in the works that has to be out of tree for a while
because it depends on new APIs for Hadoop Common. Was thinking it should go
up into GH in the meantime.
Makes sense for big features to go on a branch and then be subject to the 3
+1 rule for merge. More eyes on the change. Also makes sense for refactors
to be done in place. They may cause pain for work on a branch, but it's
less painful to merge in as it happens on the main branch than sort out
collisions later if a refactor and feature want to go in at about the same
time.
And on this part:
- A bunch of discussions on the RPC with KeyValue/Cells
.... / Tags :-)
On Tue, Feb 19, 2013 at 5:27 PM, Devaraj Das <[email protected]> wrote:
> - Stabilizing 0.96 / CI
> -- not running continuously
> -- Roman: is it a good idea to run IT under Bigtop
> - the HBase unit tests are good..
> - What about HBase as a backend for hive - tests for those
> - Do we think there is value with the integration with the rest
> of the hadoop ecosystem
> -- Jon: What's needed to make this work
> -- Roman: Collective agreement that we want to solve this
> -- Stack: We need to run the IT tests from Enis and Sergey running
> continuously
> -- Roman: If we all agree that this is something needs to be fixed
> .. then yes we would talk about mechanics. Bigtop doesn't have
> expertise in HBase and hence HBase folks would have to debug failures.
> -- Roman: Bigtop is committed to tests. Less than a dozen tests for
> hbase currently..
> -- We have been running validation around RC time. Find all kinds of
> issues - sometimes trivial (maybe a config issue),
> -- Roman can offer CI for trunk but will work for only hadoop-2 line.
> -- Roman does first line of triage.
> -- No issues to do with other ecosystem artifacts. Bigtop ensures
> the right artifacts are in place.
> -- hadoop-2 is important but not particular about the version of
> hadoop within 2. For example 2.0.2
> -- Gary: Will be good to run security tests
> -- Roman & Devaraj to talk on how this can be done/implemented
>
> On 0.96 branching
> - Lars:
> -- We will have three branches to maintain.
> - Stack: we need to stabilize quickly
> - Enis: What about 0.95.
> - Stack: Just do the snapshots thing. Every week, give a snapshot
> - Enis: we've done a bunch of stuff in RPC.. If we have to break
> something, we can if it is beta (0.96-beta).
> - Agreement is there generally ... Debate on the name with snapshot
> versus 95.0/96.0
> -- 0.95 experimental .. 0.96 will be stable
> -- If we go with 0.95, releasing will be easier as well..
> -- 0.95 will not be in production..
> -- 0.96 will be off 0.95 branch and not trunk based
>
> - How do we go about committing issues.. (issue commit rate is low)
> -- 0.94 is stable - 2 new commits and 2 new bugs a day
> -- 0.96 has lots of issues not reviewed
> -- Break up the patch into multiple smaller pieces to make review easier
> -- Branching on a big feature was suggested -
> --- Issues: committer needs to be there
> Sergey: If the rate of change is high on common code (to the
> branches) then merging will be tough
> --- Jon: Refactor should be done in the main branch (since it
> doesn't add any new funtionality)
> --- Release often to reduce #backports overall and issues with that..
>
> - Review process .. how to drive a review to closure. Effort goes
> waste if the review process is not completed. The same reviewer should
> continue to review the patch ..
> - Hard to enforce any process
> - Enis: there should be a summary of the patch and all that .. so that
> the review process is helped.. Hard to understand the architecture of
> the patch unless documented
> - Jon: It should be easy to make a one-to-one correspondence between
> the description and the patch
> - The commit should have only the jira# as opposed to pages of description
>
> - Component owners: is this working. Committers need to be forthcoming
> with reviews
> -- Maybe review the modules and add some more if needed.
>
> -- Good that we have more contributions coming than we have
> reviewers, but unless we keep up, we will plateau
> -- Mail on dev@ list if review doesn't happen
>
> - Dev co-ordination:
> -- How best can we pull together
> -- Priorities:
> --- Getting 0.96 out is priority
> --- Backports to 0.94 will happen .. until 0.96 is stable
> --- 0.92 release ? Any committer who wants to make a release can do
> so (maybe with some backports, etc.)
> --- Backporting can be tough if there are bugs and the bugfix has
> to be applied to all branches
>
> - HBASE-2600 - this requires a change in the client and the server.
> They have to be changed in lock-step. Its hard to do this .. Jon
> doesn't want to have the fix for 0.96. So 0.98 might be another
> singular release. Maybe do a rewrite of the meta after taking a lock
> on the meta, do a shim layer to handle the backward compat. between
> 0.96 and 0.98
>
> - What do people want to get into 0.94
> -- The biggest thing - Snapshots, mostly new code, about a 3rd of the
> stuff in 0.94 already
> -- Compactions improvments - no backport
>
> - How devs can better co-ordinate
> -- Snapshots co-ordination working well
> -- One page design is useful (makes it readable and all)
> -- How about handling the stripe compaction - where an idea leads to
> a bunch of others
> -- Again write-up should be done
>
> - Should we change the description to match the comments
> -- Two ideas suggested:
> --- We probably should have the description updated with the
> "Date: new description" if the issue at hand is updated
> --- Should we have a summary after a bunch of comments - yes
>
> - The face-to-face meetings are useful. We should semd out the minutes
> of the discussion to the dev list. We probably should have more
> focused huddles. Discuss but don't decide (decide on the jira)!
>
> - Jon: Would people be amenable to merge sooner rather than later on
> snapshots? Tested and being beaten up.
> - Stack: Yes
>
> - What else goes in in 0.96:
> -- RPC refactor
> -- ROOT removal
> -- Compaction stuff
> -- Package name mailing list thread - there is now a jira on that.
> We shouldn't break clients. Package name changes is not worth the
> trouble.
>
> - A bunch of discussions on the RPC with KeyValue/Cells
>
> - What do we do about usability
> -- It'll be nice if we don't need to change configs..
> -- Maybe expose more metrics and then allow for online config
> changes since automatic config is difficult and needs to be battle
> tested and all.
>
> - Benchmarking of the release:
> -- We should measure the overhead of PB stuff
>
--
Best regards,
- Andy
Problems worthy of attack prove their worth by hitting back. - Piet Hein
(via Tom White)
