Fix how we check for the unsealed status of a piece in Shard migration #575
Conversation
| if len(pinfo.Deals) <= 0 { | ||
| return xerrors.Errorf("no storage deals found for Piece %s", pcid) | ||
| } |
There was a problem hiding this comment.
length can never be below zero.
| if len(pinfo.Deals) <= 0 { | |
| return xerrors.Errorf("no storage deals found for Piece %s", pcid) | |
| } | |
| if len(pinfo.Deals) == 0 { | |
| return xerrors.Errorf("no storage deals found for Piece %s", pcid) | |
| } |
| @@ -105,6 +105,8 @@ type StorageProviderNode interface { | |||
|
|
|||
| // GetProofType gets the current seal proof type for the given miner. | |||
| GetProofType(ctx context.Context, addr address.Address, tok shared.TipSetToken) (abi.RegisteredSealProof, error) | |||
|
|
|||
| IsUnsealed(ctx context.Context, sectorID abi.SectorNumber, offset abi.UnpaddedPieceSize, length abi.UnpaddedPieceSize) (bool, error) | |||
storagemarket/testnodes/testnodes.go
Outdated
| n.lk.Lock() | ||
| defer n.lk.Unlock() | ||
|
|
||
| sealed := n.Sealed[sectorID] |
There was a problem hiding this comment.
If a sector is not tracked, this will return the zero value of the bool type which is false, and subsequently we'd return true, nil. We should fail if the map existence check fails.
(I know this is a stub for testing only, but let's fix it so that it doesn't send us down time-consuming debugging trips in the future!)
| type SectorStateAccessor interface { | ||
| StateSectorGetInfo(context.Context, address.Address, abi.SectorNumber, types.TipSetKey) (*miner.SectorOnChainInfo, error) | ||
| IsUnsealed(ctx context.Context, sector storage.SectorRef, offset storiface.UnpaddedByteIndex, size abi.UnpaddedPieceSize) (bool, error) | ||
| } | ||
|
|
There was a problem hiding this comment.
nit: I actually preferred @dirkmc's method of tightening the interface requirement to only the 2 methods that are needed. It's more idiomatic and specific (defining interface at the point of use). This new approach leaks, in sense that now the ShardMigrator is now privy to the fact that these two behaviours are provided by two different system components.
That said, no need to revert this at all, it's absolutely not a big deal here because this component has a limited lifetime and is not reusable.
Nice catch, I didn't realize this 👍 |
…fix/shard-migration-isunsealed
@dirkmc
We can have multiple deals for a Piece. We wanna know if there exists a single deal for the Piece that can be served from an Unsealed sector file.
We can't pass offset=0,size=1 to check the unsealing status because the sector storage subsystem needs that infromation to determine if we an unsealed copy of the deal (even if a sector file is already unsealed, it's possible for it to have holes i.e. it only has some unsealed deals and dosen't have any data for others).