One of the use cases that would benefit from this would be metrics sending.
Given that the current statsd protocol lacks the ability to supply
metadata, such as job and index ids, some apps have taken to inserting what
would otherwise be tagged data into the metric namespace. As an
Screenshot 2015-09-10 17.25.19.png]
Endpoints like Datadog and OpenTSDB want key names that are not unique per
instance. Graphite has wildcard semantics to accomodate this. But Datadog
and OpenTSDB do not, and would need this implemented elsewhere in the
delivery chain. StatsD doesn't provide a way to side-channel this
information, and we don't want to implement custom parsing on consumers
when we overload the metric key.
I believe that this protocol will be a move towards providing a better
means by which folks can supply metrics to the system without having to
make convention decisions that have to be scraped out and transformed on
the consumer side, as wasn't done above. A generic schema does not exist
currently, and this appears to be a promising way of delivering that
It would be much easier to use a generic schema to output to DataDog,
OpenTSDB, Graphite, and others, than it would be to guess a schema from a
flattened result (for example, "router__0" is understandably job and index,
but what does the "vizzini_1_abcd" part represent? How would I parse this
if I didn't have a human trace it back to source?).
On Tue, Sep 8, 2015 at 7:29 AM Johannes Tuchscherer <jtuchscherer(a)pivotal.io>
I guess I am working under the assumption that the current upstream schema
is not going to see a terrible amount of change. The StatsD protocol has
been very stable for over four years, so I don't understand why we would
add more and more metric types. (I already struggle with the decision to
have container metrics as their own data type. I am not quite sure why that
was done vs just expressing them as ValueMetrics).
I am also not following your argument with the multiple implementations of
a redis export? Why would you have multiple implementations of a redis info
export? Also, why does the downstream consumer have to know about the
schema? Neither the datadog nozzle nor the graphite nozzle cares about any
type of schema right now.
But to answer your question, I think as a downstream developer I am not as
interested in whether you are sending me a uint32 or uint64, but the
meaning (e.g. counter vs value) is much more important to me. So, if you
were to do nested metrics, I think I would rather like to see having nested
counters or values in there plus maybe one type that we are missing which
is a generic event with just a string.
Generally, I would try to avoid falling into the trap of creating a overly
generic system at the cost of making consumers unnecessarily complicated.
Maybe it would help if you outlined a few use cases that might benefit from
a system like this and how specifically you would implement a downstream
consumer (e.g. is there a common place where I can fetch the schema for the
generic data point?).
On Sat, Sep 5, 2015 at 6:57 AM, James Bayer <jbayer(a)pivotal.io> wrote:
after understanding ben's proposal of what i would call an extensible
generic point versus the status quo of metrics that are actually hard-coded
in software on by the metric producer and the metric consumer, i
immediately gravitated toward the approach by ben.
cloud foundry has really benefited from extensibility in these examples:
* diego lifecycles
* app buildpacks
* app docker images
* app as windows build artifact
* service brokers
* cf cli plugins
* collector plugins
* firehose nozzles
* diego route emitters
* garden backends
* bosh cli plugins
* bosh releases
* external bosh CPIs
* bosh health monitor plugins
let me know if there are other points of extension i'm missing.
in most cases, the initial implementations required cloud foundry system
components to change software to support additional extensibility, and some
of the examples above still require that and it's an issue in frustration
as someone with an idea to explore needs to persuade the cf maintaining
team to process a pull request or complete work on an area. i see ben's
proposal as making an extremely valuable additional point of extension for
creating application and system metrics that benefits the entire cloud
i am sympathetic to the question raised by dwayne around how large the
messages will be. it would seem that we could consider an upper bound on
the number of attributes supported by looking at the types of metrics that
would be expressed. the redis info point is already 84 attributes for
all of the following seem related to scaling considerations off the top
of my head:
* how large an individual metric may be
* at what rate the platform should support producers sending metrics
* what platform quality of service to provide (lossiness or not, back
pressure, rate limiting, etc)
* what type of clients to the metrics are supported and any limitations
related to that.
* whether there is tenant variability in some of the dimensions above.
for example a system metric might have a higher SLA than an app metric
should we consider putting a boundary on the "how large an individual
metric may be" by limiting the initial implementation to a number of
attributes (that we could change later or make configurable?).
i'm personally really excited about this new set of extensibility being
proposed and the creative things people will do with it. having loggregator
as a built-in system component versus a bolt-on is already such a great
capability compared with other platforms and i see investments to make it
more extensible and apply to more scenarios as making cloud foundry more
valuable and more fun to use.
On Fri, Sep 4, 2015 at 10:52 AM, Benjamin Black <bblack(a)pivotal.io>
the problem of upstream schema changes causing downstream change or
breakage is the current situation: every addition of a metric type implies
a change to the dropsonde-protocol requiring everything downstream to be
the schema concerns are similar. currently there is no schema whatsoever
beyond the very fine grained "this is a name and this is a value". this
means every implementation of redis info export, for example, can, and
almost certainly will, be different. this results in every downstream
consumer having to know every possible variant or to only support specific
variants, both exactly the problem you are looking to avoid.
i share the core concern regarding ensuring points are "sufficiently"
self describing. however, there is no clear line delineating what is
sufficient. the current proposal pushes almost everything out to schema. we
could imagine a change to the attributes that includes what an attribute is
(gauge, counter, etc), what the units are for the attribute, and so on.
it is critical that we balance the complexity of the points against
complexity of the consumers as there is no free lunch here. which specific
functionality would you want to see in the generic points to achieve the
balance you prefer?
On Wed, Sep 2, 2015 at 2:07 PM, Johannes Tuchscherer <
The current way of sending metrics as either Values or Counters through
the pipeline makes the development of a downstream consumer (=nozzle)
pretty easy. If you look at the datadog nozzle, it just takes all
ValueMetrics and Counters and sends them off to datadog. The nozzle does
not have to know anything about these metrics (e.g. their origin, name, or
Adding a new way to send metrics as a nested object would make the
downstream implementation certainly more complicated. In that case, the
nozzle developer has to know what metrics are included inside the generic
point (basically the schema of the metric) and treat each point
accordingly. For example, if I were to write a nozzle that emits metrics to
Graphite with a StatsD client (like it is done here), I need to know if
my int64 value is a Gauge or a Counter. Also, my consumer is under constant
risk of breaking when the upstream schema changes.
We are already facing this problem with the container metrics. But at
least the container metrics are in a defined format that is well documented
and not likely to change.
I agree with you, though, the the dropsonde protocol could use a
mechanism for easier extension. Having a GenericPoint and/or GenericEvent
seems like a good idea that I whole-heartedly support. I would just like to
stay away from nested metrics. I think the cost of adding more logic into
the downstream consumer (and making it harder to maintain) is not worth the
benefit of a more concise metric transport.
On Tue, Sep 1, 2015 at 5:52 PM, Benjamin Black <bblack(a)pivotal.io>
great questions, dwayne.
1) the partition key is intended to be used in a similar manner to
partitioners in distributed systems like cassandra or kafka. the specific
behavior i would like to make part of the contract is two-fold: that all
data with the same key is routed to the same partition and that all data in
a partition is FIFO (meaning no ordering guarantees beyond arrival time).
this could help with the multi-line log problem by ensuring a single
consumer will receive all lines for a given log entry in order, allowing
simple reassembly. however, the lines might be interleaved with other lines
with the same key or even other keys that happen to map to the same
partition, so the consumer does require a bit of intelligence. this was
actually one of the driving scenarios for adding the key.
2) i expect typical points to be in the hundreds of bytes to a few KB.
if we find ourselves regularly needing much larger points, especially near
that 64KB limit, i'd look to the JSON representation as the hierarchical
structure is more efficiently managed there.
On Tue, Sep 1, 2015 at 4:42 PM, <dschultz(a)pivotal.io> wrote:
I was wondering if you could give a concrete use case for the
partition key functionality.
In particular I am interested in how we solve multi line log entries.
I think it would be better to solve it by keeping all the data (the
multiple lines) together throughout the logging/metrics pipeline, but could
see how something like a partition key might help keep the data together as
Second question: how large do you see these point messages getting
(average and max)? There are still several stages of the logging/metrics
pipeline that use UDP with a standard 64K size limit.
On Aug 28, 2015, at 4:54 PM, Benjamin Black <bblack(a)pivotal.io>
The existing dropsonde protocol uses a different message type for
each event type. HttpStart, HttpStop, ContainerMetrics, and so on are all
distinct types in the protocol definition. This requires protocol changes
to introduce any new event type, making such changes very expensive. We've
been working for the past few weeks on an addition to the dropsonde
protocol to support easier future extension to new types of events and to
make it easier for users to define their own events.
The document linked below  describes a generic data point message
capable of carrying multi-dimensional, multi-metric points as sets of
name/value pairs. This new message is expected to be added as an additional
entry in the existing dropsonde protocol metric type enum. Things are now
at a point where we'd like to get feedback from the community before moving
forward with implementation.
Please contribute your thoughts on the document in whichever way you
are most comfortable: comments on the document, email here, or email
directly to me. If you comment on the document, please make sure you are
logged in so we can keep track of who is asking for what. Your views are
not just appreciated, but critical to the continued health and success of
the Cloud Foundry community. Thank you!