Make key prefixes nice
mrocklin opened this issue · 3 comments
Let's remember that keyname prefixes get used in the dashboard and are a common first impression that people have of the project. With that in mind it's important that, at least for common operations, these names are evocative and simple.
As an example, for dd.read_parquet(...) this should probably be "read_parquet" or "parquet" and not, as it is today "readparquetfsspec".
Some other concerns:
- p2pshuffle -> shuffle maybe?
- sum-tree / mean-tree -> sum/mean?
I would like names to be simple to understand at first glance and descriptive.
I'd think about having a general structure that we stick to, for example:
<logical expr>-<physical variant>-<task>
This would translate to sum-tree, shuffle-p2p-unpack.
Once we have a scheduler integration, we should be able to show progress on individual expressions instead of tasks which is probably a better representation for end users.
I don't object to a structure like that in theory. I do want to make sure that we're optimizing the experience for users rather than for ourselves. If the user really doesn't care about the physical variant, then it might not make sense to include it (unless it's really important for us when tracking down common failures (which it might be)).