Just wanted to point out the elegance of the return dict from

execute()

matching the dict passed as ~inputs~_*overrides*_, such that it provides really easy checkpoint-style processing.

# foo and bar are expensive to compute
setup = {…}
midpoint = dr.execute(["foo", "bar"], inputs=setup)

# Now we can use foo and bar as overrides
vars1 = {…}
results1 = dr.execute(["baz"], inputs=setup | vars1, overrides=midpoint)

vars2 = {…}
results2 = dr.execute(["baz"], inputs=setup | vars2, overrides=midpoint)

Nice!

This is a great point! Just checking might you want to be considering overrides instead of inputs?

I know so little, just getting acquainted. I'm unfamiliar with overrides. /me rushes off to the docs

Caching is another one to do that automatically — but what you pointed out is the exact initial design goal :)

Oh, but from testing I now see that what I wrote is actually wrong. It needs to be:

# What I had written
results1 = dr.execute(["baz"], inputs=setup | midpoint | vars1)

# What is needed
results1 = dr.execute(["baz"], inputs=setup | vars1, overrides=midpoint)

Yep that’s fair — i need to report back on the behavior here, whether overrides also counts as inputs

Mmm, checking.

Yeah so I generally agree that that is unfortunate, had I done it again in the first place I would have made it so overrides superceded inputs

I'll put it on the back burner of "might be nice PRs" to look into unifying them. Thanks!

as far as I remember: `inputs`: • is a

node_name: value

mapping •

node_name

must not match a node name defined by a function name •

node_name

must match a node name defined by a function argument name `overrides`: • is a

node_name: value

mapping •

node_name

must match a node name defined by a function name • with the

value

known before execution, this prevents the function execution, "overriding it". It then changes the remaining DAG path to execute

We could make overrides supersede inputs I think; but the idea was that they're separate things IIRC.