Re: Pulsar now using standard Apache BookKeeper

[Sijie Guo]

On Mon, Mar 5, 2018 at 11:22 AM, Joe Francis  wrote:

> taking BK dev off this..
>
> Pulsar storage has to be rock stable at all times.   I will make a  few
> suggestions to that end.
>
> Note that there is a direct conflict of interests here. For new
> installations, they have no baggage and would like the latest/greatest
> features. But for users who are already running it in production, stability
> to existing applications is almost always not welcome. No bleeding edge
> changes for the latter case.
>
> - Storage upgrades should be  less frequent.   They are not cheap
> operationally.  Rollback is a mitigation action only. Even rollback comes
> with a cost,  of degradation and disruption to service.   And
> loss/corruption of data is not an operationally recoverable event.  For any
> user that runs Pulsar in production, a BK upgrade (any storage upgrade fro
> that matter) has to be planned and executed to cover business risks.
>
> - Decoupling BK and Pulsar. There was a hard-wried dependency in Pulsar to
> a specific version of  BK till now.  A byproduct of that was a kind of
> forced
> storage stability across many Pulsar releases.  Going forward, stability is
> not automatic. My suggestion is that any release of Pulsar must be able to
> run on a range (min version - max version) of compatible BK releases and
> not just one version.  Now that Pulsar can run on the Apache BK, maybe
> relax the dependency to some extent.  That would make storage upgrade
> optional, till say some major release that mandated it.
>
> So while I can understand the need for a storage upgrade in this case now,
> the Pulsar community has to make a release plan and set a pace for storage
> upgrades. Storage system releases are quite different from stateless
> serving software.  So far, Pulsar upgrades have been easy, because more or
> less the storage was the same and stable. I don't want to be in a place
> where I cannot deploy a Pulsar release because of  the risks in a storage
> upgrade.  --- Is every release of Pulsar going to have a BK upgrade?  --
> How many times  will storage upgrade occur in a year? --- Is the storage
> upgrade optional?These are important questions that the community
> should  discuss.
>

My take on this:

I think bookkeeper release should be coupled from pulsar release. That said
not every release of Pulsar should upgrade BK, and should probably make BK
upgrade optional for Pulsar users. However that means
we might need to be looking into testing mechanisms to ensure Pulsar can
work with different BK releases, which can probably achieved with
arquillian based integration testing.

BK 4.7 and Pulsar 2.0 is kind of special in this picture, where Pulsar is
flipping from using yahoo specific release to apache release. I would
expect Pulsar would stay with 4.7 for a while after 2.0.

- Sijie




>
> Joe
>
>
>
> On Mon, Mar 5, 2018 at 1:41 AM, Ivan Kelly  wrote:
>
> > > Keep in mind the next Pulsar release needs to be based on a particular
> > BookKeeper release. An Apache project cannot release against another
> > project's SNAPSHOT since that has not been validated by their release
> > process.
> >
> > The bookkeeper 4.7.0 release process should be starting soon, so it
> > should be out before the pulsar 2.0 release process kicks off.
> >
> > -Ivan
> >
>

Re: Pulsar now using standard Apache BookKeeper

[Joe Francis]

taking BK dev off this..

Pulsar storage has to be rock stable at all times.   I will make a  few
suggestions to that end.

Note that there is a direct conflict of interests here. For new
installations, they have no baggage and would like the latest/greatest
features. But for users who are already running it in production, stability
to existing applications is almost always not welcome. No bleeding edge
changes for the latter case.

- Storage upgrades should be  less frequent.   They are not cheap
operationally.  Rollback is a mitigation action only. Even rollback comes
with a cost,  of degradation and disruption to service.   And
loss/corruption of data is not an operationally recoverable event.  For any
user that runs Pulsar in production, a BK upgrade (any storage upgrade fro
that matter) has to be planned and executed to cover business risks.

- Decoupling BK and Pulsar. There was a hard-wried dependency in Pulsar to
a specific version of  BK till now.  A byproduct of that was a kind of forced
storage stability across many Pulsar releases.  Going forward, stability is
not automatic. My suggestion is that any release of Pulsar must be able to
run on a range (min version - max version) of compatible BK releases and
not just one version.  Now that Pulsar can run on the Apache BK, maybe
relax the dependency to some extent.  That would make storage upgrade
optional, till say some major release that mandated it.

So while I can understand the need for a storage upgrade in this case now,
the Pulsar community has to make a release plan and set a pace for storage
upgrades. Storage system releases are quite different from stateless
serving software.  So far, Pulsar upgrades have been easy, because more or
less the storage was the same and stable. I don't want to be in a place
where I cannot deploy a Pulsar release because of  the risks in a storage
upgrade.  --- Is every release of Pulsar going to have a BK upgrade?  --
How many times  will storage upgrade occur in a year? --- Is the storage
upgrade optional?These are important questions that the community
should  discuss.

Joe



On Mon, Mar 5, 2018 at 1:41 AM, Ivan Kelly  wrote:

> > Keep in mind the next Pulsar release needs to be based on a particular
> BookKeeper release. An Apache project cannot release against another
> project's SNAPSHOT since that has not been validated by their release
> process.
>
> The bookkeeper 4.7.0 release process should be starting soon, so it
> should be out before the pulsar 2.0 release process kicks off.
>
> -Ivan
>

Re: Pulsar now using standard Apache BookKeeper

[Ivan Kelly]

> Keep in mind the next Pulsar release needs to be based on a particular 
> BookKeeper release. An Apache project cannot release against another 
> project's SNAPSHOT since that has not been validated by their release process.

The bookkeeper 4.7.0 release process should be starting soon, so it
should be out before the pulsar 2.0 release process kicks off.

-Ivan

Re: Pulsar now using standard Apache BookKeeper

[Dave Fisher]

Hi -

This is great. Now that this is in the two projects (which share committers) 
will need to co-ordinate releases.

Keep in mind the next Pulsar release needs to be based on a particular 
BookKeeper release. An Apache project cannot release against another project's 
SNAPSHOT since that has not been validated by their release process.

Best Regards,
Dave

> On Mar 2, 2018, at 3:23 AM, Ivan Kelly  wrote:
> 
> Great to see all these changes in!
> 
> -Ivan
> 
> On Thu, Mar 1, 2018 at 8:14 PM, Enrico Olivelli  wrote:
>> Great!
>> Congrats
>> 
>> Do you have any number to compare performances?
>> 
>> Enrico
>> 
>> Il gio 1 mar 2018, 19:06 Sijie Guo  ha scritto:
>> 
>>> Awesome work, Ivan & Jia!
>>> 
>>> Congrats Pulsar community!
>>> 
>>> - Sijie
>>> 
>>> On Thu, Mar 1, 2018 at 9:53 AM, Matteo Merli  wrote:
>>> 
 In Pulsar master branch, we have switched the BookKeeper dependency from
 the Yahoo fork to the 4.7.0-SNAPSHOT version of BookKeeper from main
>>> master
 branch.
 
 All the the changes that were in the Yahoo fork, which was based on 4.3.1
 release, have been already merged upstream.
 
 This was a big effort that took ~ 1 year to get through. There were 246
 commits to merge into a codebase that slightly changed in a 4 years
 timespan. For the curious, this is the spreadsheet we used to track the
 merging.
 https://docs.google.com/spreadsheets/d/1jAy3EfjViqNEKpCKpWiRv-
 PCZGzdjwm_PclL7Obog4Q/
 
 I would like to call out Ivan Kelly & Jia Zhai for giving a big push of
 porting many of the changes into BookKeeper and the BookKeeper community
 for being very receptive and helpful in getting this load of changes back
 into mainline.
 
 
 Matteo
 --
 Matteo Merli
 
 
>>> 
>> --
>> 
>> 
>> -- Enrico Olivelli



signature.asc
Description: Message signed with OpenPGP

Re: Pulsar now using standard Apache BookKeeper

[Ivan Kelly]

Great to see all these changes in!

-Ivan

On Thu, Mar 1, 2018 at 8:14 PM, Enrico Olivelli  wrote:
> Great!
> Congrats
>
> Do you have any number to compare performances?
>
> Enrico
>
> Il gio 1 mar 2018, 19:06 Sijie Guo  ha scritto:
>
>> Awesome work, Ivan & Jia!
>>
>> Congrats Pulsar community!
>>
>> - Sijie
>>
>> On Thu, Mar 1, 2018 at 9:53 AM, Matteo Merli  wrote:
>>
>> > In Pulsar master branch, we have switched the BookKeeper dependency from
>> > the Yahoo fork to the 4.7.0-SNAPSHOT version of BookKeeper from main
>> master
>> > branch.
>> >
>> > All the the changes that were in the Yahoo fork, which was based on 4.3.1
>> > release, have been already merged upstream.
>> >
>> > This was a big effort that took ~ 1 year to get through. There were 246
>> > commits to merge into a codebase that slightly changed in a 4 years
>> > timespan. For the curious, this is the spreadsheet we used to track the
>> > merging.
>> > https://docs.google.com/spreadsheets/d/1jAy3EfjViqNEKpCKpWiRv-
>> > PCZGzdjwm_PclL7Obog4Q/
>> >
>> > I would like to call out Ivan Kelly & Jia Zhai for giving a big push of
>> > porting many of the changes into BookKeeper and the BookKeeper community
>> > for being very receptive and helpful in getting this load of changes back
>> > into mainline.
>> >
>> >
>> > Matteo
>> > --
>> > Matteo Merli
>> > 
>> >
>>
> --
>
>
> -- Enrico Olivelli

Re: Pulsar now using standard Apache BookKeeper

[Enrico Olivelli]

Great!
Congrats

Do you have any number to compare performances?

Enrico

Il gio 1 mar 2018, 19:06 Sijie Guo  ha scritto:

> Awesome work, Ivan & Jia!
>
> Congrats Pulsar community!
>
> - Sijie
>
> On Thu, Mar 1, 2018 at 9:53 AM, Matteo Merli  wrote:
>
> > In Pulsar master branch, we have switched the BookKeeper dependency from
> > the Yahoo fork to the 4.7.0-SNAPSHOT version of BookKeeper from main
> master
> > branch.
> >
> > All the the changes that were in the Yahoo fork, which was based on 4.3.1
> > release, have been already merged upstream.
> >
> > This was a big effort that took ~ 1 year to get through. There were 246
> > commits to merge into a codebase that slightly changed in a 4 years
> > timespan. For the curious, this is the spreadsheet we used to track the
> > merging.
> > https://docs.google.com/spreadsheets/d/1jAy3EfjViqNEKpCKpWiRv-
> > PCZGzdjwm_PclL7Obog4Q/
> >
> > I would like to call out Ivan Kelly & Jia Zhai for giving a big push of
> > porting many of the changes into BookKeeper and the BookKeeper community
> > for being very receptive and helpful in getting this load of changes back
> > into mainline.
> >
> >
> > Matteo
> > --
> > Matteo Merli
> > 
> >
>
-- 


-- Enrico Olivelli

Re: Pulsar now using standard Apache BookKeeper

[Sijie Guo]

Awesome work, Ivan & Jia!

Congrats Pulsar community!

- Sijie

On Thu, Mar 1, 2018 at 9:53 AM, Matteo Merli  wrote:

> In Pulsar master branch, we have switched the BookKeeper dependency from
> the Yahoo fork to the 4.7.0-SNAPSHOT version of BookKeeper from main master
> branch.
>
> All the the changes that were in the Yahoo fork, which was based on 4.3.1
> release, have been already merged upstream.
>
> This was a big effort that took ~ 1 year to get through. There were 246
> commits to merge into a codebase that slightly changed in a 4 years
> timespan. For the curious, this is the spreadsheet we used to track the
> merging.
> https://docs.google.com/spreadsheets/d/1jAy3EfjViqNEKpCKpWiRv-
> PCZGzdjwm_PclL7Obog4Q/
>
> I would like to call out Ivan Kelly & Jia Zhai for giving a big push of
> porting many of the changes into BookKeeper and the BookKeeper community
> for being very receptive and helpful in getting this load of changes back
> into mainline.
>
>
> Matteo
> --
> Matteo Merli
> 
>

Pulsar now using standard Apache BookKeeper

[Matteo Merli]

In Pulsar master branch, we have switched the BookKeeper dependency from
the Yahoo fork to the 4.7.0-SNAPSHOT version of BookKeeper from main master
branch.

All the the changes that were in the Yahoo fork, which was based on 4.3.1
release, have been already merged upstream.

This was a big effort that took ~ 1 year to get through. There were 246
commits to merge into a codebase that slightly changed in a 4 years
timespan. For the curious, this is the spreadsheet we used to track the
merging.
https://docs.google.com/spreadsheets/d/1jAy3EfjViqNEKpCKpWiRv-PCZGzdjwm_PclL7Obog4Q/

I would like to call out Ivan Kelly & Jia Zhai for giving a big push of
porting many of the changes into BookKeeper and the BookKeeper community
for being very receptive and helpful in getting this load of changes back
into mainline.


Matteo
-- 
Matteo Merli