The BLOCK_SUMMARIES message type would differentiate between an empty response and a missing/invalid response. However, in V2, a response with empty summaries would throw a BufferUnderflowException and be treated by the caller as a null message.
This caused problems when trying to find a common block with peers that have diverged by more than 8 blocks. With V1 the caller would know to search back further (e.g. 16 blocks) but in V2 it was treated as "no response" and so the caller would give up instead of increasing the look-back threshold.
This fix will identify BLOCK_SUMMARIES_V2 messages with no content, and return an empty array of block summaries instead of a null message.
Should be enough to recover any stuck nodes, as long as they haven't diverged more than 240 blocks from the main chain.
This should hopefully fix a potential issue where peer's chain tip data becomes contaminated with other summary data, causing incorrect sync decisions.
The `start.sh` & `stop.sh` scripts have already been marked as executables in the source folder... But since we have only piped their contents, we need to set correct file permissions again.
This allows the first pass to always be served from the peer's cache of 8 summaries. This allows a maximum of 7 to be returned, because the 8th spot is needed for the parent block's signature.
Loading from the cache should speed up sync decisions, particularly when choose which peer to sync from. The greater the number of connected peers, the more significant this optimization will be. It should also reduce wasted network requests and data usage.
Adding this check prior to making a network request is a simple way to introduce the new cached summaries from BLOCK_SUMMARIES_V2 without having to rewrite a lot of the complex sync / peer comparison logic. Longer term we may want to rewrite that logic to read from the cache directly, but it doesn't make sense to introduce that level of risk at this point time, especially as the Synchronizer may be rewritten soon to prefer longer chains.
Even so, this is still quite a high risk commit so lots of testing will be needed.
The main difference here is that we now remove items from the onlineAccountsImportQueue in a batch, _after_ they have been imported. This prevents duplicates from being added to the queue in the previous time gap between them being removed and imported.
This caused large gaps with no presence data. They are removed when they expire, causing the local count to drop to zero, and the node would only start requesting them again once a peer had pushed one or more entries proactively.
Touches quite a few files because:
* Deprecate HEIGHT_V2 because it doesn't contain enough info to be fully useful during sync.
Newer peers will re-use BLOCK_SUMMARIES_V2.
* For newer peers, instead of sending / broadcasting HEIGHT_V2,
send top N block summaries instead, to avoid requests for minor reorgs.
* When responding to GET_BLOCK, and we don't actually have the requested block,
we currently send an empty BLOCK_SUMMARIES message instead of not responding,
which would cause a slow timeout in Synchronizer.
This pattern has spread to other network message response code,
so now we introduce a generic 'unknown' message type for all these cases.
* Remove PeerChainTipData class entirely and re-use BlockSummaryData instead.
* Each Peer instance used to hold PeerChainTipData - essentially single latest block summary - but now holds a List of latest block summaries.
* PeerChainTipData getter/setter methods modified for compatibility at this point in time.
* Repository methods that return BlockSummaryData (or lists of) now try to fully populate them,
including newly added block reference field.
* Re-worked Peer.canUseCommonBlockData() to be more readable
* Cherry-picked patch to Message.fromByteBuffer() to pass an empty, read-only ByteBuffer to subclass fromByteBuffer() methods, instead of null.
This allows natural use of BufferUnderflowException if a subclass tries to use read(), or hasRemaining(), etc. from an empty data-payload message.
Previously this could have caused an NPE.
It will now request online accounts every 1 minute instead of every 5 seconds, except for the first 5 minutes following a new online accounts timestamp, in which it will request every 5 seconds (referred to as the "burst" interval). It will also use the burst interval for the first 5 minutes after the node starts.
This is based on the idea that most online accounts arrive soon after a new timestamp begins, and so there is no need to request accounts so frequently after that. This should reduce data usage by a significant amount.
Once mempow is fully rolled out, the "burst" feature can be reduced or removed, since online accounts will be sent ahead of time, generally 15-30 mins prior to the new online accounts timestamp becoming active.
This closes a gap where accounts would be moved from onlineAccountsImportQueue to onlineAccountsToAdd, but not yet imported. During this time, there was nothing to stop them from being added to the import queue again, causing duplicate validations.
