Iteration over loaded values

[turbolent]

Issue To Be Solved

In Cadence, it would be useful to have a mechanism to iterate over an atree array or ordered map, which only considers slabs and values which have been previously loaded from storage.

For example, if a large array or map is loaded from storage, only its root slab is loaded, and none of the elements (child slabs) are loaded.
If only one element is accessed, then only the slabs in the tree leading to the element is loaded into memory from storage, other slabs and elements are not.

Suggested Solution

I do not have a full solution in mind, but investigated a bit how this could be achieved.

For example, in the array iterator:

• When retrieving a slab, the iterator should do so only if the slab had been previously retrieved (

  atree/array.go

  Line 2202 in bc0184e

  slab, found, err := i.storage.Retrieve(i.id)

  ). This probably needs an additional function in SlabStorage, or an additional parameter for the Retrieve function to indicate if the value should be retrieved if it has not been previously retrieved (default).
• When iterating over elements, the iterator should only load values if they have been previously loaded (

  atree/array.go

  Line 2218 in bc0184e

  element, err = i.dataSlab.elements[i.index].StoredValue(i.storage)

  ). This probably needs an alternative function in Storable, or an additional parameter for the StoredValue (like for Retrieve).

Would it be possible to add such a feature? Would such an approach work?

[fxamacker]

Hi @turbolent

What is the use case for iterating over just the loaded elements?

It'll be helpful to know more context (why needed & how used) because some edge cases come to mind:

• What if the loaded element is StorageID pointing to another slab?
• What if there is gap between loaded elements? For example, there are 3 data slabs for an array and only first and last slabs are loaded.

[turbolent]

@fxamacker Great questions!

What is the use case for iterating over just the loaded elements?

This idea / feature request is related to onflow/cadence#2460.

For example, imagine a large container (array or dictionary) is loaded from storage, and only one or a few elements are accessed. When the container is moved, we need to potentially perform some operations for elements of the container (in particular, if references where taken to elements, those need to be invalidated).
In that case we know though that it was only possible to create a reference if the element was first loaded, i.e. we can ignore elements that have not been loaded before.

What if the loaded element is StorageID pointing to another slab?

If a storage ID storable (a storable that points to another storable stored in another slab) is encountered, it should only be followed if the targeted slab was previously loaded.

For example, imagine a Cadence array of resources. The array is backed by an atree array, and the resources in it are backed by atree ordered maps, the array is only directly storing storage ID storables.
If a reference is created to the second element, only it gets loaded into memory, the other elements are not.
Thus, when iterating, only the storage ID storable for the loaded slab ID should be followed (i.e. result in a load via StoredValue, and iterator callback).

What if there is gap between loaded elements? For example, there are 3 data slabs for an array and only first and last slabs are loaded.

As far as I understand, the same logic applies for iteration/loading of internal slabs: I think only the loaded data slabs have to be considered, because by definition, a load of an element that is stored by a data slab in the "gap" should have caused a load of that data slab (and its parents in the tree).

[bluesign]

One moment when reading I got hope for lazy arrays and dictionaries. ( wishful reading )

But then I thought why not have that as it also covers this case. I spent very little time on atree but maybe this gives some ideas so I am sharing. ( It can be totally stupid, in that case sorry in advance to take your time )

What if what we have something like Hydrate method on Value. Normally everything will be dehydrated ( something like pointer with type ) but user of atree ( in this case Cadence ) hydrate values ( something like pointer dereference ) when needed.

We will be able to navigate all object tree with dehydrated values, hydrate something if we need.

In this use case Cadence can iterate and check if value is hydrated.

[turbolent]

@bluesign Great points and explanation!

atree's data structures (arrays and ordered maps) are basically already lazy. Loading an array or map from storage only loads their root node, and only once elements are accessed, those accessed elements are loaded, and nothing else.

The "dehydrated" elements are Storables, the "hydrated" elements/values are Values, the "hydrate" function is Storable.StoredValue() Value, and the "dehydrate" function is Value.Storable() Storable.

The "problem" is that the current iteration functionality operates on high-level/"hydrated" Values, instead of low-level/"dehydrated" Storables.

Offering iteration functionality on Storable level would be a good first step towards avoiding loading values / dehydration. However, what is proposed here goes one step further: it would be even more efficient to not even consider Storables in a container that were never read before / loaded into memory (what Faye is referring to above as gaps in the tree).

[fxamacker]

@bluesign Thanks for great points! I think Bastian's reply covered everything.

@turbolent Thanks for explanation and links to Cadence issue+PR for more context.

Would it be possible to add such a feature? Would such an approach work?

I think this feature is possible but we might need a different approach than suggested to iterate loaded elements.

Currently iterator traverses data slabs using next storage ID to avoid loading more metadata slabs. So it doesn't handle "gaps" of unloaded data slabs. We need a different iterator to traverse data slabs from parent slabs.

To help me prioritize, when do we need this feature?

[turbolent]

@fxamacker

I think this feature is possible but we might need a different approach than suggested to iterate loaded elements.

Currently iterator traverses data slabs using next storage ID to avoid loading more metadata slabs. So it doesn't handle "gaps" of unloaded data slabs. We need a different iterator to traverse data slabs from parent slabs.

Nice!

To help me prioritize, when do we need this feature?

It is not urgent, but would be "needed" for the Stable Cadence release, as it features reference invalidation on resource invalidation (move/deletion; see linked Cadence PR). @SupunS is looking into alternative solutions, but it feels like this would be the "most efficient", in terms of avoiding the addition of a new bookkeeping mechanism to Cadence for this purpose.

[SupunS]

Thanks @fxamacker for looking into this!

Yeah, like Bastian mentioned, I've been looking into alternatives, but apart from the one in the original PR onflow/cadence#2460 (which is inefficient), other alternatives are turning out to be mostly dead-ends.

And yes, it is not a blocker for the moment, but would be a blocker for releasing Stable Cadence. The existing PR (onflow/cadence#2460) would only need a very little change once this is available I believe (meaning that there are no big changes waiting on this feature).

Let me know if you need more context on the use-case/how this would be used, I can maybe set up a quick call to walk through an example.

[fxamacker]

Quick update. I told @j1010001 this should take about 3-5 days total (most of that for writing tests). Will try to open PR this week (most likely tomorrow if there are no surprises or emergencies because its going much faster than expected).

[turbolent]

Sounds great @fxamacker! This isn't urgent BTW, we only need it for releasing Stable Cadence, which is several months out