when making many (e.g. 1000s) of @op calls, they become increasingly slow

[amakelov]

The culprit seems to be InMemCallStorage.save, specifically these lines:

        for k, v in call.inputs.items():
            self.df.loc[(call.hid, k), :] = ("in", call.cid, v.cid, v.hid, call.op.name, call.semantic_version, call.content_version)
        for k, v in call.outputs.items():
            self.df.loc[(call.hid, k), :] = (
                "out",
                call.cid,
                v.cid,
                v.hid,
                call.op.name,
                call.semantic_version,
                call.content_version,
            )

probably because the df grows too much over time. The solution is to change the storage format for the in-memory storage.