aurora-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From 黄 凯 <>
Subject 答复: Discussion on review request 51536
Date Fri, 02 Sep 2016 23:42:45 GMT
Another concern is that once we rolled out the new executor, we should rolled out a new client
in order to use the health-check feature. Hence the executor and client rolling out process
seem to be coupled.

发件人: 黄 凯 <>
发送时间: 2016年9月3日 7:23
收件人: Zameer Manji;
抄送: Joshua Cohen;;;
主题: 答复: Discussion on review request 51536

Thanks for the new proposal, Zameer. It sounds good to me. The benefit is that it does not
alter the current infrastructure too much.

However, there is one thing to keep in mind:

we currently do a check to ensure watch_sec is longer than initial_interval_secs. We will
have to remove the alert message if we choose to skip watch_sec by setting it as zero.

So the new configuration will not support executor-driven health check unless the executors
are rolled out 100%.

Does this tradeoff seems OK for us, Maxim?


发件人: Zameer Manji <>
发送时间: 2016年9月3日 6:53
抄送: 黄 凯; Joshua Cohen;;;
主题: Re: Discussion on review request 51536

On Fri, Sep 2, 2016 at 3:24 PM, Maxim Khutornenko <<>>
Need to correct a few previous statements:

> Also we do not want to expose this message to users.
This is incorrect. The original design proposal suggested to show this
message in the UI as: "Task is healthy"

Does this mean the message in the status update is going to be exactly, "Task is healthy"
and the scheduler is going to check for this string in the `TASK_RUNNING` status update? This
means we are going to establish a communication
mechanism between the executor and scheduler that's not defined by a schema. I feel that's
worse than putting JSON in there and having the scheduler parse it.

> The Mesos API isn't designed for packing arbitrary data
> in the status update message.
Don't think I agree, this is exactly what this field is for [1] and we
already use it for other states [2].

I guess I should have said 'structured arbitrary data'. The informational, messages are fine
and we plumb them blindly into our logging and UI. I'm not convinced we should start putting
JSON or something more structured in there. That's yet another schema we have and yet another
versioning story we have to go though. This also complicates matters for custom executor authors.

> I would be open to just saying that scheduler version
> 0.16 (or 0.17) just assumes the executor transitions to
> RUNNING once a task is healthy and dropping
> `watch_secs`entirely.
We can't drop 'watch_secs' entirely as we still have to babysit job
updates that don't have health checks enabled.

Understood. I guess we can keep it but I'm now frustrated that we have a parameter that is
ignored if we set some json in Ideally, we don't accept `watch_secs`
if we want health check driven updates. As mentioned before I don't like this implicit tightening
of the executor and the scheduler.

As for my take on the above, I favor #1 as the simplest answer to an
already simple question: "Should we use watch_secs for this instance
or not?". That's pretty much it. Scheduler does not need any schema
changes, know what health checks are or if a job has them enabled. At
least not until we attempt to move to centralized health checks
(AURORA-279) but that will be an entirely different design discussion.

[1] -
[2] -

With all of this in mind, I have another proposal. Why can't we have the executor changes
(wait until the task is healthy for RUNNING) *and* read `watch_secs` if it is set? Why not
have both of these features and if users want purely health checking driven updates they can
set this value to 0 and enable health checks. If they want to have both health checking and
time driven updates they can set this to value to the time that they care about. If they just
want time driven updates they can disable health checking and set this value.

Then there is no coupling between the executor and the scheduler except for status updates
and there is no dependency on the `message` field of the status update.

We could even treat `watch_secs` as minimum time in STARTING + RUNNING instead of RUNNING
with this change and it becomes the lower bound in the update transition speed. This can ensure
that users don't deploy "too fast" and end up overwhelming other services if they are deployed
too quickly.

On Fri, Sep 2, 2016 at 2:26 PM, Zameer Manji <<>>
> *cc: Renan*
> I think there is some disagreement/discussion on the review because we have
> not achieved consensus on the design. Since the design doc was written,
> Aurora adopted multiple executor support as well non HTTP based
> healthchecking. This invalidates some parts of the original design. I think
> all of the solutions here are possible amendments to the design doc.
> I am not in favor of Solution 2 at all because status updates between
> executor <-> agent <-> master <-> scheduler are designed to update
> framework of updates to the task and not really designed to send arbitrary
> information. Just because the Mesos API provides us with a string field
> doesn't mean we should try to pack in arbitrary data. Also, it isn't clear
> what other capabilities we might add in the future so I'm unconvinced that
> capabilities needs to exist at all. My fear is that we will create the
> infrastructure for capabilities just to serve this need and nothing else.
> I object to Solution 1 along the same lines. The Mesos API isn't designed
> for packing arbitrary data in the status update message and I don't think
> we should abuse that and rely on that. Also our current infrastructure just
> plumbs the message to the UI and I think displaying capabilities is not
> something we should do.
> I am in favor of Solution 3 which is as close as possible to the original
> design in the design doc. The design doc says the following:
> Scheduler updater will skip the minWaitInInstanceMs (aka watch_secs
>> <>)
>> grace period any time it detects a named port ‘health’ in task
>> configuration. A RUNNING instance status will signify the end of instance
>> update.
> Instead of detecting the 'health' port in the task configuration, we make
> enabling this feature explicitly by enabling a bit in the task
> configuration with the `executorDrivenUpdates` bit.
> I understand this option makes this feature more complex because it
> requires a schema change and requires operators to deploy the executor to
> all agents before upgrading the client. However, I think that's a one time
> operational cost as a opposed to long lived design choices that will affect
> the code.
> Further Solution 3 is the most amenable to custom executors and continues
> our tradition of treating executors as opaque black boxes. I think there is
> a lot of value in treating executors as black boxes as it leaves the door
> open to switching our executor to something else and doesn't impose a
> burden to others that want to write their own.
> Alternatively, if amending the schema is too much work, I would be open to
> just saying that scheduler version 0.16 (or 0.17) just assumes the executor
> transitions to RUNNING once a task is healthy and dropping `watch_secs`
> entirely. We can put it in the release notes that operators must deploy the
> executor to 100% before deploying the scheduler.
> On Thu, Sep 1, 2016 at 6:40 PM, 黄 凯 <<>>
>> Hi Folks,
>> I'm currently working on a feature on aurora scheduler and executor. The
>> implementation strategy became controversial on the review board, so I was
>> wondering if I should broadcast it to more audience and initiate a
>> discussion. Please feel free to let me know your thoughts, your help is
>> greatly appreciated!
>> The high level goal of this feature is to improve reliability and
>> performance of the Aurora scheduler job updater, by relying on health check
>> status rather than watch_secs timeout when deciding an individual instance
>> update state.
>> Please see the original review request *
>> <> *
>> aurora JIRA ticket *
>> <>*
>> design doc *
>> <>*
>> for more details and background.
>> Note: The design doc becomes a little bit outdated on the "scheduler
>> change summary" part (this is what the review request trying to address).
>> As a result, I've left some comment to clarify the latest proposed
>> implementation plan for scheduler change.
>> There are two questions I'm trying to address here:
>> *1. How does the scheduler infer the executor version and be backward
>> compatible?*
>> *2. Where do we determine if health check is enabled?*
>> In short, there are 3 different solutions proposed on the review board.
>> In the first two approaches, the scheduler will rely on a string to
>> determine the executor version. We determine whether health check is
>> enabled merely on executor side. There will be communication between the
>> executor and the scheduler.
>> *Solution 1: *
>> *vCurrent executor sends a message in its health check thread during
>> RUNNING state transition, and the vCurrent updater will infer the executor
>> version from the presence of this message, and skip the watch_secs if
>> necessary.*
>> *Solution 2:*
>> *Instead of relying on the presence of an arbitrary string in the message,
>> rely on the presence of a string like:
>> "capabilities:CAPABILITY_1,CAPABILITY-2" where CAPABILITY_1 and
>> CAPABILITY_2 (etc.) are constants defined in api.thrift. Basically just
>> formalizing the mechanism and making it a bit more future proof.*
>> In the third solution, the scheduler infers the executor version from the
>> JobUpdateSettings on scheduler side.
>> *Solution 3:*
>> *Adding a bit to JobUpdateSettings which is ‘executorDrivenUpdates', if
>> that is set, the scheduler assumes that the transition from STARTING ->
>> RUNNING makes the executor healthy and concurrently, we release thermos and
>> change HealthCheckConfig to say that it should only go to running after
>> healthy*.
>> *Pros and Cons:*
>> The main benefit of Solution 1 is:
>> 1. By using the message in task status update, we don't have to make any
>> schema change, which makes the design simple.
>> 2. The feature is fully backward-compatible. When we roll out the vCurrent
>> schedulers and executors, we do not have to instruct the users to provide
>> additional field in the Job or Update configs, which could confuses
>> customers when the vPrev and vCurrent executor coexist in the cluster.
>> Concerns:
>> Relying on the presence of a message makes things brittle. Also we do not
>> want to expose this message to users.
>> The benefit of Solution 2 is making the feature more future proof.
>> However, if we do not envision a new executor feature in the short term,
>> it's not too much different from Solution 1.
>> The benefits of Solution 3 include:
>> 1. We support more than just thermos now (and others rely on custom
>> executors).
>> 2. A lot of things in Aurora treat the executor as opaque. The status
>> update message sent by executor should not be visible to users only if it's
>> an error message.
>> Concerns:
>> 1. In addition to the ‘executorDrivenUpdates' bit that identifies the
>> executor version, we still need to notify the scheduler if health check is
>> enabled on vCurrent executor, if not, the scheduler must be able to fall
>> back to use watch_secs.
>> 2. The users have to provide an additional field in their .aurora config
>> files. The feature wouldn't be available unless new clients are rolled out
>> as well.
>> Please let me know if I understand your suggestions correctly and
>> hopefully everyone is on the same page!
>> Thanks,
>> Kai

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