fix: Fix watching already mined transactions #997

popzxc · 2024-06-30T12:33:35Z

Potentially fixes #941

Changes provider and Heartbeat in a way that accounts for recently mined transactions.

Motivation

I've met a problem where .watch() would get stuck even though a transaction is mined on the node.
When I checked the code, I noticed that we only start looking for it in blocks that come after we subscribe for transaction.

Solution

Check if the transaction is already mined in the Provider trait.
- ⚠️ There is an edge case where the transaction is mined but also has custom number of confirmations. I didn't find a way to extract the block number from an abstract transaction receipt, so just left a TODO there. I hope that it's an edge case that can be tackled separately, though -- the solution here would cover all the "simple" cases.
Add a lookbehind buffer into Heartbeat, so that we can check if the transaction was confirmed recently.
Make sure that the heartbeat is spawned before we submit the first transaction.

Added test was failing before the fix, now it passes.

PR Checklist

Added Tests
Added Documentation
Breaking changes

mattsse

thanks, I believe this should work.

left some comments/qs

mattsse · 2024-07-02T10:57:50Z

crates/provider/src/heart.rs

+            }
+        }
+        self.past_blocks.push_back((*block_height, block.transactions.hashes().copied().collect()));
+        error!(?self.past_blocks, "lookbehind");


why is this an error?

Prototyping artifact, fixed.

mattsse · 2024-07-02T10:59:54Z

crates/provider/src/provider/trait.rs

+                return Ok(PendingTransaction::ready(*config.tx_hash()));
+            }
+            // TODO: There exists a corner case, if we subscribe to an already mined transaction and
+            // require a custom number of confirmations. The `TransactionReceipt` trait


this sounds like something we want to add to the trait.
we can do that separately though

Implemented via #1003, thanks.

mattsse · 2024-07-02T11:04:22Z

crates/provider/src/heart.rs

+                warn!(%block_height, last_height, "non-continuous chain");
+            }
+        }
+        self.past_blocks.push_back((*block_height, block.transactions.hashes().copied().collect()));


this doesn't handle reorgs, right?
here we'd append blocks with an existing height?
and we could find a tx confirmed in a reorged block in L473

so I think this should replace the block with same number if it exists

@mattsse That's right!

However, turns out the issue with reorgs is much deeper.
Once reorg occurs, we must also drop all the existing subscriptions received at block >= *block_height.
This part is easy, but it doesn't end here.
ChainStreamPoller doesn't seem to account for reorgs at all, e.g. here it always expects the chain to be continuous.
And to fix that we probably need to compare block tip against next_yield here, and if block_number < next_yield, we should do some kind of search for the last non-diverging block (so probably we need to poll not only for block number, but rather for block header to use root hashes for comparison).

There is also an issue with the poller stream sporadically ending on transient network error, but that's a separate problem.

I've pushed a WIP commit with reorg handling in the Heartbeat and a (very flaky due to issues above) test.
PTAL, and let me know how you want to proceed.

If it works, I suggest the following:

I will polish the commit, so that heartbeat supports reorgs, but will remove the test for now.

We will merge this PR as-is, and I will create an issue for reorg handling in ChainStreamPoller.

I will create a separate pull request for said issue, and will restore the test for reorg detection. (I cannot fully commit that I will do it due to limited capacity; but given that I already did some research and the problem seems interesting, I'll try my best to deliver it this week).

thanks,
suggestions sgtm!

I've removed the parts that are not relevant for this PR and created an issue for the remaining work.
PTAL.

mattsse

one more q re reap_reorg

mattsse · 2024-07-03T13:23:28Z

crates/provider/src/heart.rs

+    /// Reap transactions overridden by the reorg.
+    fn reap_reorg(&mut self, new_height: u64) {


what does this do?
needs some additional docs, not obvious what this does?

looks like this drops the channel? which would mean on reorg, the listeners is unable to obtain the receipt because the watcher dropped?

what does this do?

It erases all the subscriptions for the blocks that were affected by the reorg. Basically, it works the same as reap_timeouts above.

looks like this drops the channel?

It does, similarly to how it does on timeout. If the block with your tx was reorganized, you are not guaranteed to receive a receipt for it, so IMHO it makes sense to drop it.

A dedicated error type would be nice (even if just for timeout), but it's a change in public interfaces, and not sure if it belongs to this PR (I can submit it in a separate PR though).

Added more information to the method doc-comment.

so IMHO it makes sense to drop it.

I don't think this makes sense, because there's a high chance this tx will also be included in the new chain.

I see, so would putting it back to unconfirmed make sense then?

mattsse · 2024-07-03T13:25:24Z

crates/provider/src/heart.rs

+    /// Lookbehind blocks in form of mapping block number -> vector of transaction hashes.
+    past_blocks: VecDeque<(u64, Vec<B256>)>,


this should probably be a hashmap of hashes?

otherwise with >200txs per block we'll spend some time searching those vecs

You mean HashMap<u64, HashSet<B256>>?
I agree about HashSet, but for the wrapper IMHO we have a nice pop_back / push_front flow here, so probably I'd suggest making it VecDeque<(u64, HashSet<B256>)>. WDYT?

Changed to VecDeque<HashSet<(u64, HashSet<B256>)>.

mattsse · 2024-07-03T13:31:12Z

crates/provider/src/heart.rs

+            // Check that the chain is continuous.
+            if *last_height + 1 != *block_height {
+                // Remove all the transactions that were reset by the reorg.
+                warn!(%block_height, last_height, "reorg detected");
+                self.reap_reorg(*block_height);
+                // Remove past blocks that are now invalid.
+                self.past_blocks.retain(|(h, _)| h < block_height);


maybe block_height <= last_height is more accurate here?

No, block_height is the first block of the "new" chain after reorg, so transactions that belong to this block should also be removed. We will push the block with block_height right below.

crates/provider/src/heart.rs

mattsse

this should behave as expected now.

but ngl, handling all of this is more complex than it should be -.-

* fix: Fix watching already mined transactions * fix: Account for reorgs in Heart

popzxc requested review from DaniPopes, gakonst, mattsse, onbjerg, prestwich and Evalir as code owners June 30, 2024 12:33

popzxc force-pushed the popzxc-fix-watch branch from 24242ff to b9ab121 Compare June 30, 2024 14:33

mattsse requested changes Jul 2, 2024

View reviewed changes

fix: Fix watching already mined transactions

0d271d0

popzxc force-pushed the popzxc-fix-watch branch 2 times, most recently from 09a6140 to 73cd403 Compare July 3, 2024 12:59

popzxc mentioned this pull request Jul 3, 2024

[Bug] ChainStreamPoller doesn't handle reorgs #1005

Open

mattsse requested changes Jul 3, 2024

View reviewed changes

popzxc force-pushed the popzxc-fix-watch branch from 73cd403 to 7be8c5f Compare July 3, 2024 17:49

mattsse requested changes Jul 4, 2024

View reviewed changes

crates/provider/src/heart.rs Show resolved Hide resolved

fix: Account for reorgs in Heart

659c80f

popzxc force-pushed the popzxc-fix-watch branch from 7be8c5f to 659c80f Compare July 4, 2024 12:39

mattsse approved these changes Jul 8, 2024

View reviewed changes

mattsse merged commit e6b3edd into alloy-rs:main Jul 8, 2024
22 checks passed

popzxc deleted the popzxc-fix-watch branch July 8, 2024 12:36

popzxc mentioned this pull request Jul 9, 2024

[FEAT] Add examples of using era-test-node with alloy matter-labs/anvil-zksync#300

Closed

ben186 pushed a commit to ben186/alloy that referenced this pull request Jul 27, 2024

fix: Fix watching already mined transactions (alloy-rs#997)

9ce29e5

* fix: Fix watching already mined transactions * fix: Account for reorgs in Heart

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: Fix watching already mined transactions #997

fix: Fix watching already mined transactions #997

popzxc commented Jun 30, 2024

mattsse left a comment

mattsse Jul 2, 2024

popzxc Jul 3, 2024

mattsse Jul 2, 2024

popzxc Jul 3, 2024

mattsse Jul 2, 2024

popzxc Jul 3, 2024

mattsse Jul 3, 2024

popzxc Jul 3, 2024

mattsse left a comment

mattsse Jul 3, 2024

popzxc Jul 3, 2024

popzxc Jul 3, 2024

mattsse Jul 4, 2024

popzxc Jul 4, 2024

mattsse Jul 3, 2024

popzxc Jul 3, 2024

popzxc Jul 3, 2024

mattsse Jul 3, 2024

popzxc Jul 3, 2024

mattsse left a comment

		/// Reap transactions overridden by the reorg.
		fn reap_reorg(&mut self, new_height: u64) {

		/// Lookbehind blocks in form of mapping block number -> vector of transaction hashes.
		past_blocks: VecDeque<(u64, Vec<B256>)>,

fix: Fix watching already mined transactions #997

fix: Fix watching already mined transactions #997

Conversation

popzxc commented Jun 30, 2024

Motivation

Solution

PR Checklist

mattsse left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mattsse left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mattsse left a comment

Choose a reason for hiding this comment