I had to read this several times, but have some suggestions. I think when you
say "action's arguments", you mean action-configured params, e.g. `wsk action
create --param p1 v1`?
My preferences would be:
- we should split off "run" args into context and params - this is the
convention change for redefining main(args) as main(context, args) we have
discussed in the past.
- I support either having init receive action-configured params
- activation args that are possibly overridden should behave exactly as
specified args - is it important that action-configured args are actually
overridden, if the context and params are separated? (receive both values, and
logic must decide when to use which)
- let's not use env variables for any arg that is variable per activation - it
is impossible if you support concurrency, and unneeded if we pass the context
to "run".
Regarding Matt's suggestion to remove init - I like this idea, but I have
concerns compared to knative which might serve every function with a different
container, vs having some containers reused for multiple functions. In the case
where we init code into an already running container, it is useful to have the
init process separate from run, since otherwise each runtime will need to track
its own init state and queue requests during init etc. If I'm not getting the
whole picture with knative, please correct me.
Thanks
Tyson
On 6/24/19, 8:43 AM, "Rodric Rabbah" <rod...@gmail.com> wrote:
In the current activation model, an action's arguments are always provided
to the action on "run", not "init".
Should we consider partitioning the argument list into two sets, the first
is exported as environment variables at "init" time, and the second become
the action's argument at "run" time? A criteria for partitioning is that
the environment variable starts with a capital letter, which is a common
convention.
For example, an action which is invoked with a JSON object
{ "XYZ": true,
"abc" : false }
would receive {"abc": false} as its arguments and can read XYZ from the
environment (as process.env.XYZ == "true" in Node.js).
This change would:
1. require a change in the invoker to pass arguments during initialization
2. require a change in the runtime proxies to export the arguments to the
environment at initialization time (additional work may be implied by 1b)
3. an annotation on actions to opt into this partitioning for backward
compatibility or to opt out. For example '-a env-partition-arguments true'
partitions the arguments and actions without this annotation are not
affected.
Some obvious question:
Q1a. should the invoker perform the partitioning or delegate it to the
runtime? The advantage of the former is that the runtimes do not have to
implement the filtering policy and do less work. I think it makes sense to
do this invoker side for uniformity.
Q1b. should the partitioning treat environment variables as immutable post
init and ignore the partition on warm starts? This is an issue when a value
is overridden during POST invoke only since for a webaction, you cannot
override a value that's already defined (and updating a bound parameter on
an action invalidates warm containers). I think env vars should be treated
as immutable despite the issue with POST invoke.
-r
