You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@tinkerpop.apache.org by Ken Hu <ke...@gmail.com> on 2024/05/27 21:01:46 UTC
Re: [DISCUSS] Replace WebSockets with HTTP 1.1
One other thing to note with the removal of Sessions is that there will no
longer be any ability to send Groovy scripts that depend on one another.
For example:
"x = 1 + 1"
"y = x + 1"
won't work anymore because the x variable won't be saved to a Session's
bindings. The impact of this change should be reduced because we will also
be switching to gremlin-lang by default which never had this capability.
This mostly affects users that want to use gremlin-groovy to send remote
commands to the server.
On Wed, Apr 3, 2024 at 5:01 PM Ken Hu <ke...@gmail.com> wrote:
> One thing to add to this is that SASL is difficult to support with
> HTTP/1.1 so it will likely need to be dropped. This means that only Basic
> Auth will be supported with HTTP/1.1. Support for SASL can be added back
> when we move to HTTP/2. However, I'm not sure how widely SASL was used. If
> there aren't any complaints from users when it is temporarily removed for
> HTTP/1.1 then maybe we should just consider dropping support for it
> altogether and not adding it back.
>
> On Mon, Mar 4, 2024 at 3:02 PM Ken Hu <ke...@gmail.com> wrote:
>
>> I believe that I've answered most questions about this and there doesn't
>> seem to have been any objections. Therefore, I'm going to assume consensus
>> has been reached and that we will move forward with HTTP only in the next
>> major release.
>>
>> As part of the work to move to HTTP, I would like to simplify some other
>> parts of the server that I believe aren't being used. This will help to
>> simplify some of the options and arguments used in the HTTP API.
>> Specifically, I would like to remove support for OpProcessors, Sessions,
>> and the UnifiedChannelizer. OpProcessor is an interesting concept to allow
>> providers to add and modify server functionality but it doesn't seem like
>> it was ever used (other than by the default ones provided). Sessions are
>> tied to WebSocket connections and the threaded nature of the current
>> Transaction API. However, the Transaction API will get reworked and
>> WebSockets is being removed which makes the Session concept obsolete. The
>> UnifiedChannelizer, I believe, was supposed to simplify Sessions but since
>> sessions don't make much sense anymore, it probably doesn't make sense to
>> keep the UnifiedChannelizer around either.
>>
>> Is anyone aware of any implementations that are actively making use of
>> these? If not, I think it makes sense to remove or replace them in the next
>> major version as well.
>>
>> On Thu, Feb 29, 2024 at 12:52 PM Ken Hu <ke...@gmail.com> wrote:
>>
>>> Thanks for the responses so far.
>>>
>>> I think we're still open to having multiple protocols as long as there
>>> is some form of commitment to maintaining it and there is demand for it
>>> within the community. This would mean that we could add grpc but someone
>>> would have to be willing to step up to maintain it. Also, I view HTTP/1.1
>>> as a starting point for the transition from WebSockets to HTTP. The
>>> features you mentioned that would be beneficial are part of HTTP/2 which, I
>>> think, is where we probably want to head longer term.
>>>
>>> From an end user perspective, there should hopefully be little to no
>>> change at all. The underlying protocol used to communicate with the server
>>> is mostly transparent to them when they use the GLVs. The only real
>>> difference is that they will probably have to adjust some of their
>>> connection options which is something users have raised issues with before
>>> and this would improve that situation. The users that potentially will get
>>> affected the most are those that use a third party driver, they may have to
>>> continue using 3.x unless those drivers are updated. For providers, this
>>> change shouldn't affect them as long as they use the Gremlin Server (which
>>> from what I can tell includes most providers). In the end, I believe this
>>> is a change that may cause some minor differences for users, but will
>>> mostly help with the continued maintenance of the project.
>>>
>>> On Wed, Feb 28, 2024 at 4:38 PM Josh Perryman <jo...@gmail.com>
>>> wrote:
>>>
>>>> I understand the cost/complexity argument. But do we have any sense of
>>>> the
>>>> value / feature side of things? What's the implementation look like
>>>> among
>>>> providers? Is it exclusively HTTP or do some implement both, or just
>>>> Websockets? From there, what about actual usage among users? Do the
>>>> implementers see any preference on that side?
>>>>
>>>> I think this a reasonable ask, but the way it is presented is heavily
>>>> biased toward developer preferences. Do we have any information on the
>>>> user
>>>> side?
>>>>
>>>> -Josh
>>>>
>>>> On Tue, Feb 27, 2024 at 4:29 PM Ken Hu <ke...@gmail.com> wrote:
>>>>
>>>> > Hi All,
>>>> >
>>>> > I would like to start the discussion of making HTTP the only protocol
>>>> that
>>>> > Apache TinkerPop supports starting with the next major version. There
>>>> are
>>>> > several reasons I'm suggesting this. First, we already support both
>>>> > WebSockets and HTTP in Gremlin Server which causes a maintenance
>>>> burden of
>>>> > having to support both. Also, I ran some tests and a streaming
>>>> version of
>>>> > HTTP using chunked transfer encoding performs just as well as
>>>> streaming via
>>>> > WebSockets so there won't be a performance loss in terms of latency.
>>>> > Second, HTTP is significantly more popular so it will be easier for
>>>> other
>>>> > third party clients to connect to. Third, this is an opportunity to
>>>> rework
>>>> > the remote Transaction API which in its current form causes
>>>> unnecessary
>>>> > complexity in the server. This will also serve as a good opportunity
>>>> to
>>>> > make the connection options more similar among all the GLVs.
>>>> >
>>>> > I've experimented with replacing WebSockets with HTTP (both 1.1 and
>>>> 2) and
>>>> > noticed similar performance in most cases. For HTTP/1.1, there was
>>>> some
>>>> > slowdown in cases where thousands of queries with few results were
>>>> issued
>>>> > per second as lots of HTTP connections needed to be opened, but I
>>>> feel like
>>>> > this isn't a practical use case. My recommendation for now is to go
>>>> with
>>>> > HTTP/1.1 as it has all the features we need and it has mostly been
>>>> > implemented already on the server side.
>>>> >
>>>> > To ease the transition, I'm suggesting that we maintain WebSockets
>>>> for the
>>>> > foreseeable future in the 3.x line while the 4.x line is moved to use
>>>> HTTP
>>>> > exclusively. This will also give more time for us to build out
>>>> support for
>>>> > transactions over HTTP which is the only missing feature that is
>>>> supported
>>>> > when using WebSockets.
>>>> >
>>>> > I'd like to hear any thoughts regarding this.
>>>> >
>>>> > Thanks,
>>>> > Ken
>>>> >
>>>>
>>>