Currently we call it in a loop without any delays which burns cpu cycles for little value.
On my machine I see decrease of cpu usage from 12% to 6% on a synced node after applying this fix.
We have a method get which returns a Vec, it's used in cases when we can't implement FromLmdbBytes or Readable on a desired type, eg when both traits and type are external for the current crate. This approach requires an extra allocation of a Vec, this PR introduces a method which accepts desirialzer as closure, which allows to desrialize in place without an intermidiate heap allocation.
It's still possible to get a Vec if needed, just pass a Vec constructor to get_with.
We have to make an extra allocation per db get request because key generation function to_key takes Vec. Taking byte slice (AsRef<[u8]> to be precise) also simplifes the code a bit.
Currently Writable accepts trait Write as a type parameter but Readable
takes Read as a trait object, which is not symmetrical and also less performant. This PR changes Readable trait and all places where it's used
Tx pool takes some parameters as trait objects. It's not an idiomatic Rust code, in this particular case we should use generic types. Trait object makes sense when we accept in runtime different concrete types which implement the trait as a value of the same field. It's not the case here. Trait objects come with a price - instead of method dispatch in compile time we have to accept runtime dispatch. My guess we did it to not clutter the code with type parameters, which is understandable but still suboptimal.
Currently we pass a Vec. This requires an extra allocation and copy of all elements if a caller doesn't have a Vec already, which is at least 95% of cases.
Another, a smaller issue, we have a function util::to_hex and some structs implement to_hex() on top of it, so we have a mix of it in the code. This PR introduces a trait and a blanket impl for AsRef<[u8]> which brings a uniform API (obj.to_hex()). One unfortunate case is arrays of size bigger than 32 - Rust doesn't implement AsRef for them so it requires an ugly hack (&array[..]).to_hex().
* Refactor SyncState
Method sync_error() retrun type was simplified.
update_txhashset_download() was made type safe, which eliminates a runtime enum variant's check, added an atomic status update
We allocate 17 vectors (in the heap) per 512 headers, this PR reduces the number to 1 by resuing the buffer for headers and eliminating the need in the vector of indices
* Node API: don't error on missing output
* Propagate errors from get_output*
* Rename is_unspent and small refactor
* Forgot to rename function in tests
* Change Batch get_output_pos_height type signature
* no need to rehash with index to compare output with input spending it
* compare output identifier when checking is_unspent()
* output identifier from cleanup
* first pass at rewind_single_block
and reworking rewind to simply iterate over blocks, rewinding each incrementally
* commit
* commit
* cleanup
* add test coverage for output_pos index transactional semantics during rewind
* commit
* do not store commitments in spent_index
just use the order of the inputs in the block
* compare key with commitment when cleaning output_pos index
* remove unused OutputPos struct
David Burkett