Circling back to nomenclature of tensor sharing policy

[zhoonit]

Discussion continuing about the nomenclature about policies of tensor sharing before working further.

I have a new idea coming along the discussion There are three level of guarantees a layer is making (analogical to https://en.wikipedia.org/wiki/Exception_safety), just like noexcept specifier, user setting a policy to tensor spec means user is guaranteeing a data integrity to some extent so that the memory utilization can be enhanced.

1. Identity integrity guarantee = identity will be exactly same as identity is shared
2. Data integrity guarantee = underlying data won't be changed
3. Space integrity guarantee = underlying space can be reused explictly

policy	A1(=unmanaged)	A2	B	C	D
Identity(meta-data) integrity guarantee	x	x	o	x	x
Data Integrity guarantee	x	x	o	o	x
Space integrity guarantee	x	x	o	o	o
---------------------------	-	-	-	-	-
option 1 (current)	placeholder	unique	shared	read_only_view	maybe_modifying_view
option 2	-	new	link	view	shared
option 3	placeholder	unique	shared	view	try_reuse

[taos-ci]

:octocat: cibot: Thank you for posting issue #1736. The person in charge will reply soon.

[zhoonit]

opinion about option 2: Policy B is direct interpretation of createOrExtend(), i think the term link lacks of the create sections.

Advocates about unique / shared (at option 1, 3): this will be analogical to unique_ptr,shared_ptr, which this tensor will have pretty similar ownership semantics.

[kparichay]

shared_ptr allows wrapping a small section of the underlying data, which is representing the case D. In case B, as far as I understand, as the tensor metadata has to remain same, the size must also be same.

Regarding link assumes that case B requires the two tensors remain exactly identical (like symlink). createOrExtend() is a wrapper which will resolve go to case A or case B.