Within MakeScalingHandler() there is a call to GetReplicas() which was not returning an error when a non-200 http response was received from /system/function/. The call would also return a populated struct, so the perception was that a function existed an had been scaled to zero. This meant that the function would be added to the function cache and the code would continue into SetReplicas() where an attempt would be made to scale up a non-existent function.
This change amends GetReplicas() so that it will return an error if the gateway returns anything other than a 200 reponse code from the /system/function/ endpoint. This causes MakeScalingHandler() to return earlier with an error indicating that the function could not be found. The cache.Set call is also moved to after the error check so that the cache is only updated to include existent functions.
During investigations as to the cause of #876 tests were added to function_cache to check that Get() is behaving as intended when function exists and when not. Tests are also added to plugin/external to test that GetReplicas() and SetReplicas() are following their intended modes of operation when 200 and non-200 responses are received from the gateway.
Signed-off-by: Richard Gee <richard@technologee.co.uk>
Existing code has been used for scaling up and querying replicas.
This meant the new code was deleted and there is less duplication
now.
The cache store a whole query response rather than just the
available replica count and the tests were updated. This has been
tested with Docker swarm and the image:
openfaas/gateway:scale-17-07-2018
This feature now needs the env-var of scale_from_zero to be enabled
in order to turn on the scaling behaviour.
Signed-off-by: Alex Ellis (VMware) <alexellis2@gmail.com>
This change allows functions to be "idled" or scaled to zero
replicas and then be invoked later on. There is a penalty to
scaling up - the API gateway proxy will block until the function
is ready.
A cache is included to off-set the calls to upstream API to check
on readiness along with unit tests.
Testing via scaling to zero replicas and then invoking function.
On Swarm I observed 3 seconds on an Intel Nuc i5 for scaling back
from zero replicas.
Signed-off-by: Alex Ellis (VMware) <alexellis2@gmail.com>