Hazelcast 3.6 / bosh release

[poblin-orange]

Hello @youngm We are still using the memcache-release, nice work !

By the way, we wanted to evaluate native hazelcast features, and tested the recent Hazelcast 3.6 version, including discovery SPI. Here is our first try : https://github.com/Orange-OpenSource/hazelcast-boshrelease

Still experimental (broker not yet operational), but the hazelcast auto discovery with consul is OK, and we added the management center.

Do you plan to update memcache release with latest hazelcast version?

Do you think it would make sense to join work, and maybe reuse our bosh release as part of memcache release ? Thks for your feed back Pierre

[youngm]

@poblin-orange Glad the solution it is working out for you.

Upgrading hazelcast: I have a list of 10-15 things I plan to address this year sometime. I should probably make issues for them so you are aware of them. I was tentatively planning to start that work once hazelcast 3.7 is released skipping 3.6. Simply because 3.5 is working for us and deploying a release based on 3.6 would reboot the cache.

Exposing Native Hazelcast: A year ago I probably would have said no. However, I've since learned that the memcache client landscape isn't as nice on some platforms as it is in Java. :) So, for some of my users a native hazelcast service might be a nice option.

Joining work: I'm totally open to the idea of attempting to join work. I could potentially change the memcache adapter from embedding hazelcast to connecting to a locally running hazelcast instance instead. I'll have to stew on the idea. Let me know if you have any other specific ideas.

[poblin-orange]

Thks Upgrading hazelcast: I still have to check the bosh canary mechanism on our hazelcast bosh release. Probably a drain mechanism, and single in flight policy could keep the cluster up while upgrading (provided Hazelcast peer to peer protocol is compatible between 3.6 et 3.6+).

Concerning join work, we are currently toying with the idea of composing bosh releases (a bit like cf-release). We started with java-openjdk, and try to maximise use case and potential reuse of the bosh release we produce for our own requirements. A starting point could be that i add all the hazelcast configuration properties you use. Then we could try a spike composing our 2 releases.

[youngm]

@poblin-orange

Upgrading hazelcast: My understanding is that Hazelcast doesn't maintain protocol compatibility between minor releases only patch releases [0]. So I'm not sure how going from 3.5 to 3.6 would behave.

Joining work: That sounds like a good place to start. I'm probably still a couple of months out from having time to work on this. In the meantime you are also welcome to spike on changing the memcache-hazelcast code to not embed hazelcast using a client instead. I haven't done any research regarding the implications of that yet.

[0] https://github.com/hazelcast/hazelcast/issues/14

[poblin-orange]

@youngm good. ll keep you informed if i make some progress.

[youngm]

I've upgraded hazelcast to 3.7. However, I still haven't reworded the code work with the official hazelcast bosh release making the memcache server code a client of already running hazelcast. We had some urgent fixes we had to make and I didn't have time to fully analyse a re-architecture of the system. I'll keep considering that for some future work. Until then I'll publish a new final release soon that uses Hazelcast 3.7 and other fixes.

[poblin-orange]

Ok nice work. Ill try to keep up with the hazelcast bosh release.