aurora-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From David McLaughlin <>
Subject Re: [RFC] REST / thrift & AURORA-987
Date Tue, 01 Dec 2015 17:37:31 GMT
Shouldn't we start with the design of the API itself? Won't that influence
many of the answers to these questions?

E.g. if you're just looking to port the Thrift API 1:1 to a JSON + HTTP
interface then that's a very different set of requirements to starting
fresh and doing a better job with our API.

Personally I don't think the existing Thrift API is a very good base to
build an API on top off. A lot of the endpoints are fit for one purpose
(e.g. a specific UI view or client function) rather than being flexible. I
can't tell you how many times we wanted to go in and improve the UI in some
way only to find the existing API does not give us access to the data we

So yeah, I feel like the API should be more generic with regards to data
access. So fewer, more-powerful endpoints that support complex queries.

On Mon, Nov 30, 2015 at 12:16 PM, John Sirois <> wrote:

> I’ve experimenting on for
> the past few weeks and I’d like to ask for feedback on the direction I’d
> like to head. If you’re interested in the evolution of the Aurora REST api,
> read on.
> ------------------------------
> AURORA-987 aims to create a first-class REST-like scheduler interface. I’ve
> re-familiarized myself with the codebase and come to the conclusion that
> transitioning to a 1st class REST api requires maintaining the core thrift
> API as the 1st class API until the point the REST API is fully established
> and clients can all be transitioned.
> I think this conclusion is probably uncontroversial, but the key factors
> pushing this way are:
>    1.
>    The thrift API has both wide and deep dependencies inside the Aurora
>    codebase - 276 imports across 97 files:
>    $ git grep "import org.apache.aurora.gen" -- src/main/java/ | grep
> -v "import" | wc -l
>    276
>    $ git grep "import org.apache.aurora.gen" -- src/main/java/ | grep
> -v "import" | cut -d: -f1 | sort -u | wc
> -l
>    97
>    2.
>    The thrift API is stored long-term in the log in serialized form.
> Both 1 & 2 dictate that the thrift API, at least its enums, structs and
> unions, must be maintained for the forseeable future.
> We also have the RPC API (thrift services) - which is currently a ~thin,
> but not insignificant, container of API processing logic. For example, see
> here
> <
> >
> .
> As such it seems to me the REST API should call into the existing thrift
> API to provide a stable transition and confidence in core logic of API
> method implementations.
> This leads to the following ideas for paths forward:
>    1. Hand construct a REST forwarding layer and maintain it in tandem with
>    thrift API changes.
>    2. Automate 1 such that thrift API changes cause REST API changes
>    automatically.
> The hand construction path has the obvious maintenance issues, but is
> otherwise straight-forward. The maintenance issues should not be
> overstated, since good tests and some extra review vigilance could be
> enough to make the approach work for the period of time both APIs are
> supported.
> That said, an automated solution with a single source of truth for the API
> definition is clearly preferrable given the automation is free.
> The automation is far from free though and so I’ve started investigating
> one approach to this automation to flesh out the scope.
> We already do our own thrift codegen
> <
> >
> via a custom gradle ThriftEntitiesPlugin
> <
> >
> that works around Apache thrift’s java codegen in order to generate
> immutable wrapper entities for the storage system.
> I propose taking this further and generating our own thrift API and
> entities in 1 pass through our thrift files. These would be immutable
> thrift entities 1st class with builders for modification and the entities
> and the generated service interfaces would carry extra metadata in the form
> of annotations to bind REST services and their metadata with.
> There are 2 paths I’ve considered towards this end:
>    1. Modify Apache thrift to support immutable-style java output with
>    support for thrift annotations.
>    2. Write our own thrift parser and code generator to do said same.
> I’ve been pursuing option 2 even though it sounds worse on its face. The
> swift <> project from Facebook brings
> options 1 and 2 back on the same level of undertaking since the parsing and
> protocol implementations can be leveraged as libraries and only the codegen
> portion need be undertaken (You can see some of that work here
> <
> >
> ).
> So, with that background 2 questions of the same form:
>    1. Is there some other fundamental approach I’m missing to bolting on a
>    1st class REST API, or is the hand construction approach favorable?
>    2. Is the approach to single point of API control using swift misguided?
>    Should I be focusing on Apache thrift enhancement instead? Should I be
>    generating the *.thrift files instead from a new 1st class source of
> truth
>    in the form of a json api schema?
> Any and all feedback is welcome!
> ​

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message