I'm in complete agreement with all the points Matt made. I think the way forward should be to expose ALL user-modifiable configs via Ambari and let Ambari actively manage them. We should keep the command line tools as the backend and Ambari should continue to leverage them. This will allow manual installation/management if desired and will ensure the command line scripts are kept up to date.
Fully leveraging Ambari has many beneficial effects. My top four: Provides proper revision control for the configurations Scales easily into things like rolling|quick upgrades and Kerberos support Provides other applications a restful endpoint to change configurations We get a force multiplier from the Ambari devs The working description Matt provided is completely consistent with my understanding of how it works (derived from the Ambari docs, authoring pieces of the mpack and interacting with some Ambari devs). Restarting Ambari agent is the only circumstance I'm aware of outside of save/start|restart that would initiate a re-write of the configs and cache, there could be others. -D... On Thu, Jan 12, 2017 at 9:24 PM, Matt Foley <ma...@apache.org> wrote: > Mike, could you try again on the image, please, making sure it is a simple > format (gif, png, or jpeg)? It got munched, at least in my viewer. Thanks. > > Casey, responding to some of the questions you raised: > > I’m going to make a rather strong statement: We already have a service > “to intermediate and handle config update/retrieval”. > Furthermore, it: > - Correctly handles the problems of distributed services running on > multi-node clusters. (That’s a HARD problem, people, and we shouldn’t try > to reinvent the wheel.) > - Correctly handles Kerberos security. (That’s kinda hard too, or at least > a lot of work.) > - It does automatic versioning of configurations, and allows viewing, > comparing, and reverting historical configs > - It has a capable REST API for all those things. > It doesn’t natively integrate Zookeeper storage of configs, but there is a > natural place to specify copy to/from Zookeeper for the files desired. > > It is Ambari. And we should commit to it, rather than try to re-create > such features. > Because it has a good REST API, it is perfectly feasible to implement > Stellar functions that call it. > GUI configuration tools can also use the Ambari APIs, or better yet be > integrated in an “Ambari View”. (Eg, see the “Yarn Capacity Scheduler > Configuration Tool” example in the Ambari documentation, under “Using > Ambari Views”.) > > Arguments are: Parsimony, Sufficiency, Not reinventing the wheel, and Not > spending weeks and weeks of developer time over the next year reinventing > the wheel while getting details wrong multiple times… > > Okay, off soapbox. > > Casey asked what the config update behavior of Ambari is, and how it will > interact with changes made from outside Ambari. > The following is from my experience working with the Ambari Mpack for > Metron. I am not otherwise an Ambari expert, so tomorrow I’ll get it > reviewed by an Ambari development engineer. > > Ambari-server runs on one node, and Ambari-agent runs on each of all the > nodes. > Ambari-server has a private set of py, xml, and template files, which > together are used both to generate the Ambari configuration GUI, with > defaults, and to generate configuration files (of any needed filetype) for > the various Stack components. > Ambari-server also has a database where it stores the schema related to > these files, so even if you reach in and edit Ambari’s files, it will Error > out if the set of parameters or parameter names changes. The historical > information about configuration changes is also stored in the db. > For each component (and in the case of Metron, for each topology), there > is a python file which controls the logic for these actions, among others: > - Install > - Start / stop / restart / status > - Configure > > It is actually up to this python code (which we wrote for the Metron > Mpack) what happens in each of these API calls. But the current code, and > I believe this is typical of Ambari-managed components, performs a > “Configure” action whenever you press the “Save” button after changing a > component config in Ambari, and also on each Install and Start or Restart. > > The Configure action consists of approximately the following sequence (see > disclaimer above :-) > - Recreate the generated config files, using the template files and the > actual configuration most recently set in Ambari > o Note this is also under the control of python code that we wrote, and > this is the appropriate place to push to ZK if desired. > - Propagate those config files to each Ambari-agent, with a command to set > them locally > - The ambari-agents on each node receive the files and write them to the > specified locations on local storage > > Ambari-server then whines that the updated services should be restarted, > but does not initiate that action itself (unless of course the initiating > action was a Start command from the administrator). > > Make sense? It’s all quite straightforward in concept, there’s just an > awful lot of stuff wrapped around that to make it all go smoothly and > handle the problems when it doesn’t. > > There’s additional complexity in that the Ambari-agent also caches (on > each node) both the template files and COMPILED forms of the python files > (.pyc) involved in transforming them. The pyc files incorporate some > amount of additional info regarding parameter values, but I’m not sure of > the form. I don’t think that changes the above in any practical way unless > you’re trying to cheat Ambari by reaching in and editing its files > directly. In that case, you also need to whack the pyc files (on each > node) to force the data to be reloaded from Ambari-server. Best solution > is don’t cheat. > > Also, there may be circumstances under which the Ambari-agent will detect > changes and re-write the latest version it knows of the config files, even > without a Save or Start action at the Ambari-server. I’m not sure of this > and need to check with Ambari developers. It may no longer happen, altho > I’m pretty sure change detection/reversion was a feature of early versions > of Ambari. > > Hope this helps, > --Matt > > ================================================ > From: Michael Miklavcic <michael.miklav...@gmail.com> > Reply-To: "dev@metron.incubator.apache.org" <dev@metron.incubator.apache. > org> > Date: Thursday, January 12, 2017 at 3:59 PM > To: "dev@metron.incubator.apache.org" <dev@metron.incubator.apache.org> > Subject: Re: [DISCUSS] Ambari Metron Configuration Management consequences > and call to action > > Hi Casey, > > Thanks for starting this thread. I believe you are correct in your > assessment of the 4 options for updating configs in Metron. When using more > than one of these options we can get into a split-brain scenario. A basic > example is updating the global config on disk and using the > zk_load_configs.sh. Later, if a user decides to restart Ambari, the cached > version stored by Ambari (it's in the MySQL or other database backing > Ambari) will be written out to disk in the defined config directory, and > subsequently loaded using the zk_load_configs.sh under the hood. Any global > configuration modified outside of Ambari will be lost at this point. This > is obviously undesirable, but I also like the purpose and utility exposed > by the multiple config management interfaces we currently have available. I > also agree that a service would be best. > > For reference, here's my understanding of the current configuration > loading mechanisms and their deps. > > <image> > > Mike > > > On Thu, Jan 12, 2017 at 3:08 PM, Casey Stella <ceste...@gmail.com> wrote: > > In the course of discussion on the PR for METRON-652 > <https://github.com/apache/incubator-metron/pull/415> something that I > should definitely have understood better came to light and I thought that > it was worth bringing to the attention of the community to get > clarification/discuss is just how we manage configs. > > Currently (assuming the management UI that Ryan Merriman submitted) configs > are managed/adjusted via a couple of different mechanism. > > - zk_load_utils.sh: pushed and pulled from disk to zookeeper > - Stellar REPL: pushed and pulled via the CONFIG_GET/CONFIG_PUT > functions > - Ambari: initialized via the zk_load_utils script and then some of them > are managed directly (global config) and some indirectly > (sensor-specific > configs). > - NOTE: Upon service restart, it may or may not overwrite changes on > disk or on zookeeper. *Can someone more knowledgeable than me about > this describe precisely the semantics that we can expect on > service restart > for Ambari? What gets overwritten on disk and what gets updated > in ambari?* > - The Management UI: manages some of the configs. *RYAN: Which configs > do we support here and which don't we support here?* > > As you can see, we have a mishmash of mechanisms to update and manage the > configuration for Metron in zookeeper. In the beginning the approach was > just to edit configs on disk and push/pull them via zk_load_utils. Configs > could be historically managed using source control, etc. As we got more > and more components managing the configs, we haven't taken care that they > they all work with each other in an expected way (I believe these are > true..correct me if I'm wrong): > > - If configs are modified in the management UI or the Stellar REPL and > someone forgets to pull the configs from zookeeper to disk, before they > do > a push via zk_load_utils, they will clobber the configs in zookeeper > with > old configs. > - If the global config is changed on disk and the ambari service > restarts, it'll get reset with the original global config. > - *Ryan, in the management UI, if someone changes the zookeeper configs > from outside, are those configs reflected immediately in the UI?* > > > It seems to me that we have a couple of options here: > > - A service to intermediate and handle config update/retrieval and > tracking historical changes so these different mechanisms can use a > common > component for config management/tracking and refactor the existing > mechanisms to use that service > - Standardize on exactly one component to manage the configs and regress > the others (that's a verb, right? nicer than delete.) > > I happen to like the service approach, myself, but I wanted to put it up > for discussion and hopefully someone will volunteer to design such a thing. > > To frame the debate, I want us to keep in mind a couple of things that may > or may not be relevant to the discussion: > > - We will eventually be moving to support kerberos so there should at > least be a path to use kerberos for any solution IMO > - There is value in each of the different mechanisms in place now. If > there weren't, then they wouldn't have been created. Before we try to > make > this a "there can be only one" argument, I'd like to hear very good > arguments. > > Finally, I'd appreciate if some people might answer the questions I have in > bold there. Hopefully this discussion, if nothing else happens, will > result in fodder for proper documentation of the ins and outs of each of > the components bulleted above. > > Best, > > Casey > > > > > >
