* fn: allow specified docker networks in functions
If FN_DOCKER_NETWORK is specified with a list of
networks, then agent driver picks the least used
network to place functions on.
* add mutex comment
* fn: experimental prefork recycle and other improvements
*) Recycle and do not use same pool container again option.
*) Two state processing: initializing versus ready (start-kill).
*) Ready state is exempt from rate limiter.
* fn: experimental prefork pool multiple network support
In order to exceed 1023 container (bridge port) limit, add
multiple networks:
for i in fn-net1 fn-net2 fn-net3 fn-net4
do
docker network create $i
done
to Docker startup, (eg. dind preentry.sh), then provide this
to prefork pool using:
export FN_EXPERIMENTAL_PREFORK_NETWORKS="fn-net1 fn-net2 fn-net3 fn-net4"
which should be able to spawn 1023 * 4 containers.
* fn: fixup tests for cfg move
* fn: add ipc and pid namespaces into prefork pooling
* fn: revert ipc and pid namespaces for now
Pid/Ipc opens up the function container to pause container.
Added env FN_MAX_FS_SIZE_MB, which if defined and non-zero
is passed to docker as storage opt size. We do not validate
if this option is supported by docker currently. This is
because it's difficult to actually validate this since it
not only depends on storage driver and its backing filesystem,
but also the mount options used to mount that fs.
* fn: enable failing test back
* fn: fortifying the stderr output
Modified limitWriter to discard excess data instead
of returning error, this is to allow stderr/stdout
pipes flowing to avoid head-of-line blocking or
data corruption in container stdout/stderr output stream.
*) I/O protocol parse issues should shutdown the container as the container
goes to inconsistent state between calls. (eg. next call may receive previous
calls left overs.)
*) Move ghost read/write code into io_utils in common.
*) Clean unused error from docker Wait()
*) We can catch one case in JSON, if there's remaining unparsed data in
decoder buffer, we can shut the container
*) stdout/stderr when container is not handling a request are now blocked if freezer is also enabled.
*) if a fatal err is set for slot, we do not requeue it and proceed to shutdown
*) added a test function for a few cases with freezer strict behavior
* update vendor directory, add go.opencensus.io
* update imports
* oops
* s/opentracing/opencensus/ & remove prometheus / zipkin stuff & remove old stats
* the dep train rides again
* fix gin build
* deps from last guy
* start in on the agent metrics
* she builds
* remove tags for now, cardinality error is fussing. subscribe instead of register
* update to patched version of opencensus to proceed for now TODO switch to a release
* meh
fix imports
* println debug the bad boys
* lace it with the tags
* update deps again
* fix all inconsistent cardinality errors
* add our own logger
* fix init
* fix oom measure
* remove bugged removal code
* fix s3 measures
* fix prom handler nil
we have been getting these from attach all this time and never needed these
anyway.
I ran cpu profiles of dockerd and this was 90% of docker cpu usage (json
logs). woot. this will reduce i/o quite a bit, and we don't have to worry
about them taking up any disk space either.
from tests i get about 50% speedup with these off. the hunt continues...
previously we would retry infinitely up to the context with some backoff in
between. for hot functions, since we don't set any dead line on pulling or
creating the image, this means it would retry forever without making any
progress if e.g. the registry is inaccessable or any other temporary error
that isn't actually temporary. this adds a hard cap of 10 retries, which
gives approximately 13s if the ops take no time, still respecting the context
deadline enclosed.
the case where this was coming up is now tested for and was otherwise
confusing for users to debug, now it spits out an ECONNREFUSED with the
address of the registry, which should help users debug without having to poke
around fn logs (though I don't like this as an excuse, not all users will be
operators at some point in the near future, and this one makes sense)
closes#727
closes#317
we could fiddle with this, but we need to at least bound these. this
accomplishes that. 1m is picked since that's our default max log size for the
time being per call, it also takes a little time to generate that many bytes
through logs, typically (i.e. without trying to). I tested with 0, which
spiked the i/o rate on my machine because it's constantly deleting the json
log file. I also tested with 1k and it was similar (for a task that generated
about 1k in logs quickly) -- in testing, this halved my throughput, whereas
using 1m did not change the throughput at all. trying the 'none' driver and
'syslog' driver weren't great, 'none' turns off all stderr and 'syslog' blocks
every log line (boo). anyway, this option seems to have no affect on the
output we get in 'attach', which is what we really care about (i.e. docker is
not logically capping this, just swapping out the log file).
using 1m for this, e.g. if we have 500 hot containers on a machine we have
potentially half a gig of worthless logs laying around. we don't need the
docker logs laying around at all really, but short of writing a storage driver
ourselves there don't seem to be too many better options. open to idears, but
this is likely to hold us over for some time.
* so it begins
* add clarification to /dequeue, change response to list to future proof
* Specify that runner endpoints are also under /v1
* Add a flag to choose operation mode (node type).
This is specified using the `FN_NODE_TYPE` environment variable. The
default is the existing behaviour, where the server supports all
operations (full API plus asynchronous and synchronous runners).
The additional modes are:
* API - the full API is available, but no functions are executed by the
node. Async calls are placed into a message queue, and synchronous
calls are not supported (invoking them results in an API error).
* Runner - only the invocation/route API is present. Asynchronous and
synchronous invocation requests are supported, but asynchronous
requests are placed onto the message queue, so might be handled by
another runner.
* Add agent type and checks on Submit
* Sketch of a factored out data access abstraction for api/runner agents
* Fix tests, adding node/agent types to constructors
* Add tests for full, API, and runner server modes.
* Added atomic UpdateCall to datastore
* adds in server side endpoints
* Made ServerNodeType public because tests use it
* Made ServerNodeType public because tests use it
* fix test build
* add hybrid runner client
pretty simple go api client that covers surface area needed for hybrid,
returning structs from models that the agent can use directly. not exactly
sure where to put this, so put it in `/clients/hybrid` but maybe we should
make `/api/runner/client` or something and shove it in there. want to get
integration tests set up and use the real endpoints next and then wrap this up
in the DataAccessLayer stuff.
* gracefully handles errors from fn
* handles backoff & retry on 500s
* will add to existing spans for debuggo action
* minor fixes
* meh
* add per call stats field as histogram
this will add a histogram of up to 240 data points of call data, produced
every second, stored at the end of a call invocation in the db. the same
metrics are also still shipped to prometheus (prometheus has the
not-potentially-reduced version). for the API reference, see the updates to
the swagger spec, this is just added onto the get call endpoint.
this does not add any extra db calls and the field for stats in call is a json
blob, which is easily modified to add / omit future fields. this is just
tacked on to the call we're making to InsertCall, and expect this to add very
little overhead; we are bounding the set to be relatively small, planning to
clean out the db of calls periodically, functions will generally be short, and
the same code used at a previous firm did not cause a notable db size increase
with production workload that is worse, wrt histogram size (I checked). the
code changes are really small aside from changing to strfmt.DateTime,
adding a migration and implementing sql.Valuer; needed to slightly modify the
swap function so that we can safely read `call.Stats` field to upload at end.
with the full histogram in hand, we can compute max/min/average/median/growth
rate/bernoulli distributions/whatever very easily in a UI or tooling. in
particular, this data is easily chartable [for a UI], which is beneficial.
* adds swagger spec of api update to calls endpoint
* adds migration for call.stats field
* adds call.stats field to sql queries
* change swapping of hot logger to exec, so we know that call.Stats is no
longer being modified after `exec` [in call.End]
* throws out docker stats between function invocations in hot functions (no
call to store them on, we could change this later for debug; they're in prom)
* tested in tests and API
closes#19
* add format of ints to swag
before returning the cookie in the driver, wait for health checks
https://docs.docker.com/engine/reference/builder/#healthcheck if provided.
for images that don't have health checks, this will have no affect (an added
call to inspect container, for hot it's small potatoes).
this will be useful for containers so that they can pull large files or do
setup that takes a while before accepting tasks. since this is before start,
it won't run into the idle timeout. we could likely use these for hot
containers in general and check between runs or something, but didn't do that
here.
one nascient concern is that for hot if the containers never become healthy
I don't think we will ever kill them and the slot will 'leak'. this is true
for this and for other cases (pulling image) I think, we should probably
recycle hot containers every hour or something which would also close this.
anyway, not a huge blocker I don't think, there will likely be 1 user of this
feature for a bit, it's not documented since we're not sure we want to support
it.
closes#336
* Docker stats to Prometheus
* Fix compilation error in docker_test
* Refactor docker driver Run function to wait for the container to have stopped before stopping the colleciton of statistics
* Fix go fmt errors
* Updates to sending docker stats to Prometheus
* remove new test TestWritResultImpl because we changes to support multiple waiters have been removed
* Update docker.Run to use channels not contextrs to shut down stats collector
go vet caught some nifty bugs. so fixed those here, and also made it so that
we vet everything from now on since the robots seem to do a better job of
vetting than we have managed to.
also adds gofmt check to circle. could move this to the test.sh script (didn't
want a script calling a script, because $reasons) and it's nice and isolated
in its own little land as it is. side note, changed the script so it runs in
100ms instead of 3s, i think find is a lot faster than go list.
attempted some minor cleanup of various scripts
currently:
* container ran out of memory (code 137)
* container exited with other code != 0
* unable to pull image (auth/404)
there may be others but this is a good start (the most common). notably, for
both hot and cold these should bubble up (if deterministic, which hub isn't
always), and these are useful for users to use in debugging why things aren't
working.
added tests to make sure that these behaviors are working.
also changed the behavior such that when the container exits we return a 502
instead of a 503, just to be able to distinguish the fact that fn is working
as expected but the container is acting funky (400 is weird here, so idk).
removed references to old IsUserVisible crap and slightly changed the
interface for RunResult for plumbing reasons (to get the error type,
specifically).
fixed an issue where if ~/.docker/config.json exists sometimes pulling images
wouldn't work deterministically (should be more inline w/ expectations now)
closes#275
* fix docker build
this is trivially incorrect since glide doesn't actually provide reproducible
builds. the idea is to build with the deps that we have checked into git, so
that we actually know what code is executing so that we might debug it...
all for multi stage build instead of what we had, but adding the glide step is
wrong. i added a loud warning so as to discourage this behavior in the future.
* hang the runner, agent=new sheriff
tl;dr agent is now runner, with a hopefully saner api
the general idea is get rid of all the various 'task' structs now, change our
terminology to only be 'calls' now, push a lot of the http construction of a
call into the agent, allow calls to mutate their state around their execution
easily and to simplify the number of code paths, channels and context timeouts
in something [hopefully] easy to understand.
this introduces the idea of 'slots' which are either hot or cold and are
separate from reserving memory (memory is denominated in 'tokens' now).
a 'slot' is essentially a container that is ready for execution of a call, be
it hot or cold (it just means different things based on hotness). taking a
look into Submit should make these relatively easy to grok.
sorry, things were pretty broken especially wrt timings. I tried to keep good
notes (maybe too good), to highlight stuff so that we don't make the same
mistakes again (history repeating itself blah blah quote). even now, there is
lots of work to do :)
I encourage just reading the agent.go code, Submit is really simple and
there's a description of how the whole thing works at the head of the file
(after TODOs). call.go contains code for constructing calls, as well as Start
/ End (small atm). I did some amount of code massaging to try to make things
simple / straightforward / fit reasonable mental model, but as always am open
to critique (the more negative the better) as I'm just one guy and wth do i
know...
-----------------------------------------------------------------------------
below enumerates a number of changes as briefly as possible (heh..):
models.Call all the things
removes models.Task as models.Call is now what it previously was.
models.FnCall is now rid of in favor of models.Call, despite the datastore
only storing a few fields of it [for now]. we should probably store entire
calls in the db, since app & route configurations can change at any given
moment, it would be nice to see the parameters of each call (costs db space,
obviously).
this removes the endpoints for getting & deleting messages, we were just
looping back to localhost to call the MQ (wtf? this was for iron integration i
think) and just calls the MQ.
changes the name of the FnLog to LogStore, confusing cause there's also a
`FuncLogger` which uses the Logstore (punting). removes other `Fn` prefixed
structs (redundant naming convention).
removes some unused and/or weird structs (IDStatus, CompleteTime)
updates the swagger
makes the db methods consistent to use 'Call' nomenclature.
remove runner nuisances:
* push down registry stuff to docker driver
* remove Environment / Stats stuff of yore
* remove unused writers (now in FuncLogger)
* remove 2 of the task types, old hot stuff, runner, etc
fixes ram available calculation on startup to not always be 300GB (helps a lot
on a laptop!)
format for DOCKER_AUTH env now is not a list but a map (there are no docs,
would prefer to get rid of this altogether anyway). the ~/.docker/cfg expected
format is unchanged.
removes arbitrary task queue, if a machine is out of ram we can probably just
time out without queueing... (can open separate discussion) in any case the
old one didn't really account well for hot tasks, it just lined everyone up in
the task queue if there wasn't a place to run hot and then timed them out
[even if a slot became free].
removes HEADER_ prefixing on any headers in the request to a invoke a call.
(this was inconsistent with cli for test anyway)
removes TASK_ID header sent in to hot only (this is a dupe of FN_CALL_ID,
which has not been removed)
now user functions can reply directly to the client. this means that for
cold containers if they write to stdout it will send a 200 + headers. for
hot containers, the user can reply directly to the client from the container,
i.e. with its preferred status code / headers (vs. always getting a 200).
the dispatch itself is a little http specific atm, i think we can add an
interchange format but the current version is easily extended to add json for
now, separate discussion. this eliminates a lot of the request/response
rewriting and buffering we were doing (yey). now Dispatch ONLY does input and
output, vs. managing the call timeout and having access to a call's fields.
cache is pushed down into agent now instead of in the front end, would like to
push it down to the datastore actually but it's here for now anyway. cache
delete functions removed (b/c fn is distributed anyway?). added app caching,
should help with latency.
in general, a lot of server/runner.go got pushed down into the agent. i think
it will be useful in testing to be able to construct calls without having to
invoke http handlers + async also needs to construct calls without a handler.
safe shutdown actually works now for everything (leaked / didn't wait on
certain things before)
now we're waiting for hot slots to open up while we're attempting to get ram
to launch a container if we didn't find any hot slots to run the call in
immediately. we can change this policy really easily now (no more channel
jungle; still some channels). also looking for somewhere else to go while the
container is launching now. slots now get sent _out_ of a container, vs.
a container receiving calls, which makes this kind of policy easier to
implement. this fixes a number of bugs around things like trying to execute
calls against containers that have not and may never start and trying to
launch a bazillion containers when there are no free containers. the driver api
underwent some changes to make this possible (relatively minimal, added Wait).
the easiest way to think about this is that allocating ram has moved 'up'
instead of just wrapping launching containers, so that we can select on a
channel trying to find ram.
not dispatching hot calls to containers that died anymore either...
the timeout is now started at the beginning of Submit, rather than Dispatch or
the container itself having to manage the call timeout, which was an
inaccurate way of doing things since finding a slot / allocating ram / pulling
image can all take a non-trivial (timeout amount, even!) amount of time. this
makes for much more reasonable response times from fn under load, there's
still a little TODO about handling cold+timeout container removal response
times but it's much improved.
if call.Start is called with < call.timeout/2 time left, then the call will
not be executed and return a timeout. we can discuss. this makes async play
_a lot_ nicer, specifically. for large timeouts / 2 makes less sense.
env is no longer getting upper cased (admittedly, this can look a little weird
now). our whole route.Config/app.Config/env/headers stuff probably deserves a
whole discussion...
sync output no longer has the call id in json if there's an error / timeout.
we could add this back to signify that it's _us_ writing these but this was
out of place. FN_CALL_ID is still shipped out to get the id for sync calls,
and async [server] output remains unchanged.
async logs are now an entire raw http request (so that a user can write a 400
or something from their hot async container)
async hot now 'just works'
cold sync calls can now reply to the client before container removal, which
shaves a lot of latency off of those (still eat start). still need to figure
out async removal if timeout or something.
-----------------------------------------------------------------------------
i've located a number of bugs that were generally inherited, and also added
a number of TODOs in the head of the agent.go file according to robustness we
probably need to add. this is at least at parity with the previous
implementation, to my knowledge (hopefully/likely a good bit ahead). I can
memorialize these to github quickly enough, not that anybody searches before
adding bugs anyway (sigh).
the big thing to work on next imo is async being a lot more robust,
specifically to survive fn server failures / network issues.
thanks for review (gulp)
