Skip to main content

2020-05-18 Incident Postmortem

· 4 min read

What Happened?

We activated witness refresh at block 339044 (see audit 23), which was intended to gradually clear witness tables over the course of about a week. Unfortunately we neglected to apply this refresh to the lagging ledger in the same way. Once we realized this, we issued a deactivation of the new feature at block 339291. Unfortunately it was too late and about 37 ledger entries were now divergent between the leading and lagging ledger. This led to disagreements on the block contents between the consensus members and caused a chain halt.


All times approximate, in PDT.

11:20AM Monday, May 18th: Initial chain variable transaction absorbed.
3:10PM Monday, May 18th: Initial signs of trouble as some nodes start to have issues absorbing blocks.
3:23PM Monday, May 18th: Chain halts for the first time.
5:40PM Monday, May 18th: Second chain variable transaction, disabling witness refresh (see audit 24).
6:30PM Monday, May 18th: Chain halts again, is left down.
6:51PM Monday, May 18th: Issue diagnosed.
7PM to late: Work on recovery code, eventually giving up for the night.
7:15AM Tuesday, May 19th: Issue fully understood, fix code written.
8:00AM Tuesday, May 19th: Hard reset plan formulated.
11:00AM Tuesday, May 19th: Fixed image goes GA.
11:45AM Tuesday, May 19th: Ancillary services fully updated, chain restarted, hard reset chain variable issued.
1:45PM Tuesday, May 19th: Witness reset interval chain variable submitted, returning the value to approximately one week.

Why did it happen?

The core of the problem was disagreement between the two versions of the ledger that we keep on disk. The leading ledger is the current ledger. The lagging ledger is 50 blocks behind, so that we have a stable base to play transactions into to validate things at heights between these ledgers. Problems started occurring when different members of the consensus group would disagree about which transactions they considered as valid. We use witnesses to generate the paths that PoC requests can go along, and when the witnesses on the ledger disagree, some nodes can consider the paths generated by others impossible. When, after some time, the disagreements happened between too many members, we could no longer construct a block.

Our lagging ledger had two sources of drift:

  1. The witness, PoC entry, and state channel garbage collection processes were not run on the lagging ledger.
  2. When the ledger was played forward via ledger_at, for the purposes of transaction validation, the GC processes were not invoked either, so even if the lagging ledger was correct, the played forward ledger that we would have arrived at would not be.

Garbage collection in this context means that we have rules for the removal of stale data from the ledgers that are invoked potentially at every block. Currently we have collectors for PoC challenge data, state channel data, and Hotspot challenge witness data.

How did we fix it?

There are additional complications, but these facts suffice to contextualize the fix, which was to apply the garbage collection process both during the playforward and also as the lagging ledger absorbed blocks.

Operationally, things were a bit more complicated. The team stayed up late into the night trying to understand how the ledger could reliably be repaired, but in the end, we decided that since witness data was now meant to be ephemeral, we could simply set the witness reset interval to 1, which would reset all witnesses at each block. If we let this run for 50+ blocks, longer than the gap between the leading and the lagging ledger, the witnesses should be consistent (which is to say, absent) in both versions of the ledger and we could let them start to accumulate again. This required building a new image with the new code, and then starting up the group again, skipping past the split block it had stalled on, with only the new chain var transaction in their buffers (see audit 25). Once the group was moving again, we monitored it closely for halts and issues until the ledger was consistent across almost all nodes, at which point we could submit another transaction that would allow witnesses to accumulate (see audit-26).