Varun Srinivasan on Warpcast

Content pfp

0 reply

0 recast

0 reaction

Varun Srinivasan pfp

Varun Srinivasan

Please upgrade your hubs to v1.13.4 We've fixed the issue with delayed messages and plan to min version this release. Hubs that are not upgraded by 9am PT tomorrow will disconnect from the network.

14 replies

11 recasts

72 reactions

Varun Srinivasan pfp

Varun Srinivasan

Sync has been degraded for a few days, and most hubs are missing some messages. The network will automatically heal over the next week as hubs sync with each other and fill the gaps. We're also working with neynar and airstack to trigger faster syncs with them to speed this up. If you are running a hub and affected by this, DM e.

1 reply

1 recast

22 reactions

Varun Srinivasan pfp

Varun Srinivasan

The root cause was an expensive iteration. When a hub gets a message from a peer, it iterates over its entire peer store to track some stats. As the number of peers grew the iteration took longer and the number of iterations needed kept increasing. Eventually, the iteration couldn't be completed before the next one was triggered. This caused hubs to start crashing slowly. The fix was effectively a 1-line change that moved the iteration out of the critical path.

11 replies

3 recasts

37 reactions

Varun Srinivasan pfp

Varun Srinivasan

Huge shoutout to everyone who worked on this @sanjay @sds @wazzymandias.eth @cassie

3 replies

0 recast

33 reactions

Hector pfp

How many peers do you think until the previous implementation started lagging ?

1 reply

0 recast

0 reaction

Jacob pfp

0 reply

0 recast

0 reaction

Tim Mustafin pfp

Thanks for sharing the root cause. Actually, looking at the metrics post-factum, it makes perfect sense there was some software-level limitation (CPU & Network saturation look quite flat)

0 reply

0 recast

0 reaction

BradGao pfp

Per as I know, iterating ALL peers is not a good idea in a peer-to-peer protocol. The iteration count should have a max value, no matter how many peers existing in the network

0 reply

0 recast

1 reaction

wake pfp

0 reply

0 recast

1 reaction

K pfp

If you do not need 'exact' no of peers you can let it run async. `statsd().gauge("peer_store_count", this.gossipNode.peerStoreCount());` * No "await" so it would be running in async * Releases current thread to work on other async tasks * You can still get peer count without obstructing main flow Assuming * Does not run every ms or second, this operation runs hourly / on peer joining, Otherwise a running count will be a good idea * Number of peers are not very important * resolves to accurate value

0 reply

0 recast

0 reaction

SchmRdty.eth 🎩 pfp

SchmRdty.eth 🎩

🤔 I need more… something.

0 reply

0 recast

0 reaction

joshisdead.eth pfp

@joshisdead.eth

Thanks for the update, V! And thanks to everyone who helped 🤩🥰

0 reply

0 recast

0 reaction

Penguin pfp

@sleepypenguin.eth

Thank you and appreciate the work @v!! Question though on delayed tips. If it was delayed and just synced now, will it be deducted on today’s allowance rather that the day it was posted?

0 reply

0 recast

0 reaction

Todd pfp

0 reply

0 recast

0 reaction