broker: parent-uri attribute is set for instance with no parent

[grondo]

I'm not certain, but this seems wrong:

$ flux start flux start flux getattr parent-uri
local:///var/tmp/grondo/flux-j6S7sS/local-0

It appears that the broker sets parent-uri whenever FLUX_URI is set in the environment, even if the enclosing instance isn't technically the parent of the new instance.

Is this expected behavior?

[garlick]

I think that is the expected behavior but maybe because I didn't think it through?

I guess a test instance should always be the "top" flux even if it was started by Flux? What is the fallout from the current behavior?

[grondo]

In this case flux did not start the test instance, it just happened to be started under an existing instance. It seems like an instance should only be considered the parent if it allocated resources to the child instance?

The fallout is that existence of the parent-uri attribute alone doesn't indicate that the current instance was launched by Flux. I guess both parent-uri and jobid need to be considered?

[grondo]

I also admit I haven't thought this through either, though. It is simple enough to check for a jobid attribute, and this issue was only opened to get clarification.

[garlick]

There is also this, which works as expected.

$ flux start flux start flux getattr instance-level
0

Yeah seems like when flux is not run as a flux job, all three of these should be true

• instance-level is 0
• jobid is not set
• parent-uri is not set

[grondo]

I just went through the code and currently all places where parent-uri is potentially used, the jobid attribute is also checked and both must be set before an attempt to contact the parent is made. So, currently there is no fallout.

There is this comment in src/modules/resource/inventory.c

flux-core/src/modules/resource/inventory.c

Lines 40 to 46 in b0c565f

	* Case 2 (method=job-info)
	* ------------------------
	* On the rank 0 broker, if the 'parent-uri' broker attribute is defined,
	* a connection is made to the parent broker, and R is read from the
	* job-info module. This R was assigned to the instance by the enclosing
	* instance scheduler, and includes ranks representing brokers in the
	* enclosing instance.

which probably needs a very slight tweak.

[garlick]

Does this seem right then?

diff --git a/src/broker/broker.c b/src/broker/broker.c
index 7bdb38bdd..c040f1cf9 100644
--- a/src/broker/broker.c
+++ b/src/broker/broker.c
@@ -629,7 +629,14 @@ static void init_attrs (attr_t *attrs, pid_t pid, struct flux_msg_cred *cred)
 {
     const char *val;
 
-    val = getenv ("FLUX_URI");
+    /* Set the parent-uri attribute IFF this instance was run as a job
+     * in the enclosing instance.  "parent" in this context reflects
+     * a hierarchy of resource allocation.
+     */
+    if (getenv ("FLUX_JOB_ID"))
+        val = getenv ("FLUX_URI");
+    else
+        val = NULL;
     if (attr_add (attrs, "parent-uri", val, ATTR_IMMUTABLE) < 0)
         log_err_exit ("setattr parent-uri");
     unsetenv ("FLUX_URI");

[grondo]

That looks correct to me.