kvs: improve API for namespaces

[garlick]

Now that we have the ns:NAMESPACE/ prefix for keys as well as flux_kvs_set_namespace(), do we need to support the FLUX_KVS_NAMESPACE environment variable?

(This was briefly noted in #1432)

After #1429 wreck sets FLUX_JOB_KVSPATH for sub-jobs. It seems like a utility, like the job shell, meant to operate on the per-job namespace. would read this and call flux_kvs_set_namespace() during its initialization? In the new execution system I was envisioning that FLUX_KVS_NAMESPACE would become something like ns:job-42/. or similar.

Just asking the question whether we need to continue supporting a special environment variable for the namespace.

In fact, maybe we should replace flux_kvs_set_namespace() wtih flux_kvs_chdir() or similar, to set a default key prefix that could optionally include both a namespace and a key part?

[chu11]

Generally speaking, I like the idea of removing FLUX_KVS_NAMESPACE if its truly not needed. I think it can get tricky with the prioritization of namespace specifications (i.e. flux_kvs_set_namespace() overrides environment variable, prefix overrides flux_kvs_set_namespace()). So removing one of those would make things better.

[garlick]

Let's see how #1460 shakes out and decide which needs to go, the environment variable or flux_kvs_set_namespace().

The advantage of an environment variable is we can set it for a job and then the user can run flux kvs ... in a script to manipulate the per-job KVS space.

I'm not sure I like the flux_kvs_set_namespace() call in real world use. In the job module, where we are dealing with multiple user namespaces, it gets confusing when chaining KVS calls with continuations. They key prefix seems more intuitive in that situation.

[chu11]

Was thinking about this issue given prototyping in #1860.

flux_kvs_set_namespace() works by setting an aux value in the flux_t handle. flux_kvs_get_namespace() works by checking for the aux value OR checking for the environment variable. The aux value for a namespace takes precedence over the environment variable. If neither are set, we use "primary".

If we were a user implementing flux_kvs_copy(), it'd be really ugly setting an alternate namespace for the src and/or dst locations. We'd have to determine if an aux is set or if the environment variable is set, save the old value of the namespace set in the handle, and set it back at the end. This is ugly by itself, but add in the fact there is a chained continuation, and we have multiple error paths, and there is a lookup / commit that may have different namespaces, and it starts to be ugly.

[chu11]

closing this as solved via #1977