Re: C++ APIs

[Michael Davidsaver <mdavidsaver@gmail.com>]

Sinisa,

On 07/19/2018 03:33 PM, Siniša Veseli wrote:
> Hi,
> 
> Judging from Michael's comments, documentation, epics meeting slides, etc. it looks like these new APIs are indeed meant to replace what we had since v4.5 or so. While I am not necessarily opposed to such change, I do have a few points I'd like to make:
> 
> 1) EPICS in itself is probably not the easiest software to learn and use. Having two sets of C++ APIs does not help and can be terribly confusing. If I were new to Argonne, and someone told me to write an EPICS C++ client for the first time, and then I were to discover that in epics base we have client.h and pvaClient.h that seem to be completely different, that we have two sets of examples, different documentation, and in some case different behavior between the clients, I would ask myself "What is going on here, and who the heck designed all of this...".

Can you name even one part of the EPICS ecosystem which hasn't had multiple implementations, APIs, and/or wrappers?

How many driver frameworks are there?
How many libca wrappers?
How many outright independent re-implementation of CA?
How many OPI systems?
How many archive engines?
How many alarm managers?

I think you're trying to hold me to an ideal which has never existed in this community.
Mostly I think this is a good thing.  EPICS is a toolkit, not a product line.
Of course anything can be taken to excess, but I think I would have to go quite a bit
further before I could be seriously accused of this.

And just to pick a nit.  Your count is one short.  The ChannelProvider API is just as
much as much a supported public API as either pvaClient or my new entrant.  There is
already quite a lot of confusion and a decades worth of outdated examples floating around.

> 2) Maintaining and supporting two sets of high-level APIs is not feasible in the long run. Resources that we have for epics development are scarce, and those should be used wisely. This is something that anyone sitting on the council should be concerned with. Why would we have Marty or anyone else spend lots of his time on pvaClient and pvDatabase if we are going to get rid of it in 6 months or in a year? I would personally rather spend my time on something that will have a lasting impact on the project, rather than fixing a bug or two in something that is viewed as temporary.

You want permanent software?  I'm not sure I can help you with this :)

I too would rather be spending time on other things.  Squashing bugs in PVA,
or working on a driver for a neat new piece of hardware.  However, I've found myself
spending more time fighting awkward and error prone APIs than doing productive
work.  I've been into the guts of this code base, and a few similar ones, so I know
what it's really doing, and I know it doesn't have to be this hard!

Others have reach the same conclusion.  However, instead of fighting to improve the
situation, they mostly written PVA off and moved on.  imo. this is part of the reason
this "new" code is already a decade old without seeing widespread use.  This isn't how
I work.  I'm willing to work to incrementally improve existing code.  Like EPICS Base,
and now the PVA modules.  It isn't glamorous.  It doesn't publish papers.  And when
it goes well, no one notices.

Rather than questioning why I've written some (rather thin) wrappers to try to
rationalize PVA, I wish you would ask me why I haven't just tried to rewrite
the whole thing?  Or why I haven't gone off and started again from scratch?


> 3) I would hope that EPICS development project has some sort of coherent strategy in terms of what gets developed and why, what are the recommended user interfaces, etc. Module owners have a wide latitude and control in terms of how they want things to be done. However, in a case where someone feels that a new high-level C++ API is needed because the old one is just not good, or if they wanted to make API changes that affect users or other downstream modules, I imagine there would be some group discussions and planning around that, so that we all know what is going on. We may not all agree about particular set of changes, but at least everyone would get their say and the group would build some sort of consensus around the direction that the project is taking.
> 
> It seems like there was not much discussion in this particular case. New interfaces appeared in the base code at some point, and are finding their way slowly through the rest of the system, so now we have two sets of APIs, and that's just how it is... Depending on who you talk to first, you might go this way or the other for developing your own code, and hope that the path you've chosen will not be deprecated in the next release. This just does not seem to be a good thing as far as EPICS users are concerned.

Ok, you have me here.  I didn't solicit input from a certain committee.  Several
of the individuals, yes.  But I only presented the results.  I had reasons for
avoiding this.  Some I'm not so proud of.  So here is your reward for reading this
far.  Some dirty laundry.  It isn't a secret that I've totally failed to figure out
how to work with Marty.  That relationship when wrong the first time we met, and has
continued to go wrong for 8 years.  I've been unable to find a way to approach
him about changes to his code without his getting defensive, and the discussion
degenerating into an unproductive stalemate.  This has been repeated many times
over the years, and I'm tired of it.  So rather than pick yet another fight with
Marty, I made something which works for me.  From my prospective it's already
paid off in time saved.  I think it will be useful for others, which is why I
documented it, but only time and bug reports will tell.

Still, I would have you note that at this point, I've spent the majority of my
career working with Marty's code.  In Base, asyn, and PVA.  Some I like, and
some I don't.  There are certainly aspects I wish were done differently, but
on balance these are good platforms to work from.  This is why I was willing
to step up as maintainer, first for Base and now for PVA.  I just wish Marty
could see past my criticism of PVA, and consider what it means that I care
enough to try to change it.

> I am not arguing to maintain some sort of status quo here, but I am not sure I completely agree with how are changes being introduced at the moment. If new client/server APIs are going to be significantly better, if we need to change the way we use PV requests in our code, etc., that would all be fine with me, as long as this was discussed, consequences were understood, transition plan for the old code was developed, etc.
> 
> This email is not meant to offend anyone, if I did so, then I am sorry and I apologize in advance.
> 
> I do appreciate Michael's clarifications on what happened and how things are at the moment.
> 
> Sinisa
> 
> 
> On 7/19/18 1:42 PM, Marty Kraimer wrote:
>>
>>
>> On 07/18/2018 08:05 PM, Michael Davidsaver wrote:
>>> On 07/18/2018 01:20 PM, Siniša Veseli wrote:
>>>> Hi,
>>>>
>>>> I realize I have not been attending the working group meetings lately, so I apologize in advance if I missed something.
>>>>
>>>> Reading through some of the github issues that seem to be getting a lot of attention recently, it seems like we are getting a completely new set of PVA client/server C++ APIs, which raises several questions in my mind:
>>>>
>>>> 1) Are those APIs becoming the "official" EPICS PVA client/server APIs? In other words, if I am starting a new C++ project based on EPICS7, should I start with the new interfaces, or work with the ones available in the most recent 7.0.1.1 release?
>>> My new client API (pva/client.h) is actually in the 7.0.1.1 release.  I wrote this last summer.  It is the result of my
>>> experiences using the ChannelProvider API with the PVA gateway prototype, the P4P client,
>>> the CLI tools, and in unittest code for QSRV and the PVA gateway.
>>>
>>> pva/client.h is my attempt to distill out the code common to all of these use cases.
>>>
>>> Since 7.0.1.1 I've successfully used it in the new PVA link support, some of the CLI tools, and recently
>>> rewritten the P4P client around it.  In doing this I didn't have to make any incompatible changes, just to
>>> add a few missing pieces and make bug fixes.  So from this point I expect it will remain stable.  I plan to
>>> apply it to the PVA gateway the next time I iterate.
>>>
>>> As for the (very new) server API.  This does not attempt to be fully general purpose, eg. it won't be
>>> enough to build a gateway.  Instead, it is a direct decedent of the server code common to QSRV and the unittests
>>> of the PVA gateway prototype, which is will replace.  This common code is what has held these two otherwise
>>> unrelated projects in the same source repository.
>>>
>>> So far it's only been use behind the P4P server API.  I'll consider it stable once it has been successfully
>>> used a few more times (by myself, or others).  So I suppose you could call it a "technology preview" in the
>>> upcoming 7.0.2 release (a step I think should be mandatory for all new APIs).
>>>
>>> As for "official"-ness; these APIs will be released, and I will continue to use them in my own projects,
>>> and as a tool to test the underlying PVA client and server code (hopefully finding bugs before you do).
>>>
>>> Should you use them?
>>>
>>> I'd like to get feedback from people like yourself who also have experience with both the low-level
>>> ChannelProvider API, as well as Marty's pvaClientCPP/pvDatabaseCPP.  My goals are that these new APIs should be easier to use,
>>> harder to break, and perform just as well.  If I'm not meeting these goals, then I'll do something about it.
>>>
>>>> 2) If #1 is true, what becomes of the current pvaClientCPP and pvDatabaseCPP APIs? Will those be deprecated? Or are we going to have two separate sets of high-level C++ APIs?
>>> As you allude to in #3, Marty and I disagree about pvaClientCPP and pvDatabaseCPP, of which I am not a fan.
>>> I don't know what his plans are, but I have not, and will not, dissuade him from continuing to develop these modules.
>>>
>>> So yes, it looks like there will be two sets.
>>>
>>>> 3) It looks like there are some rather significant differences in behavior between the new and existing interfaces. Have those differences been summarized anywhere?
>>> No.  I can only refer you to my documentation.  See "Client API" and "Server API", and particularly the "API usage Examples".
>>>
>>> http://epics-base.github.io/pvAccessCPP/
>>>
>>>> I would imagine that making API changes like this, and for a project that numerous sites depend on, requires some sort of process: somebody comes up with a proposal and plan, there is a discussion about advantages/disadvantages and about the consequences for the existing code and for the future development, and (most importantly) there is an agreement between the developers as far as the direction that project will be taking. If this was discussed in one of the v4 meetings already, sorry that I missed it, and I would appreciate if someone points me to the document where all of this has been written down.
>>> My motivation for this tool smithing is that I found the existing options quite awkward, error prone,
>>> and unlikely to change.  I know that others have similar experiences.  I developed some patterns and
>>> utilities to cope with this.  I also dislike copy+pasting code between projects.  These new APIs are
>>> the end result of this process.  I know they are useful to me. I think they will be useful for others,
>>> but only time will tell.
>>>
>>>> On a personal note, few years back, after we all decided that the new client API (pvaClientCPP) will be implemented, I had to put in a fair amount of work to replace my own high-level API with the official one, and I would not be exactly looking forward to yet another iteration of the same. However, if this was already discussed and agreed upon, then so be it.
>>> If you're happy with pvaClientCPP, then I see no immediate reason for you to change.  The low level
>>> ChannelProvider API on which all of these high level APIs are based will be maintained (warts and all)
>>> for the foreseeable future.
>>>
>>> Eventually, I hope that you will conclude that these new APIs would make your job easier.  But only you
>>> can be the judge of this.
>>>
>> Michael is creating a new client  API for pvAccess.
>>
>>> https://github.com/epics-base/pvAccessCPP/blob/master/src/client/pva/client.h
>>
>> He has used this API as the basis for
>> 1) pvTools. In particular pvget and pvput
>> 2) pva2pva. Both qsrv and gateway
>> 3) p4p
>>
>> The issues
>>> https://github.com/epics-base/pvAccessCPP/issues/116
>>
>> and
>>
>>> https://github.com/epics-base/pva2pva/issues/21
>>
>> Show that pvTools and qsrv have broken the exceptions of existing clients and servers.
>>
>> In order to support both the existing clients and Michael's new client API then there will have to be two implementation of at least pvTools and qsrv. If we continue with two client API's there will probably be a need for more dual implementations.
>>
>> Marty
>