Skip to content
This repository has been archived by the owner on Nov 6, 2020. It is now read-only.

Fails to sync ancient blocks from peer with the same difficulty #9225

Closed
ascjones opened this issue Jul 26, 2018 · 0 comments
Closed

Fails to sync ancient blocks from peer with the same difficulty #9225

ascjones opened this issue Jul 26, 2018 · 0 comments
Labels
F2-bug 🐞 The client fails to follow expected behavior. M4-core ⛓ Core client code / Rust. P5-sometimesoon 🌲 Issue is worth doing soon.
Milestone

Comments

@ascjones
Copy link
Contributor

I'm running:

  • Which Parity version?: 2.0.0
  • Which operating system?: Linux
  • How installed?: binaries
  • Are you fully synchronized?: yes
  • Which network are you connected to?: ethereum
  • Did you try to restart the node?: yes

Reproduce:

Create a reserved-peers file with a node you know is fast and up to date with the best block.

parity -l sync=trace --log-file=log-file --reserved-peers=reserved-peers --reserved-only

Logs like the following should be produced:

1751-2018-07-26 15:51:39  IO Worker #3 TRACE sync  Considering peer 0, force=false, td=Some(0x12f966affeebce9f577), our td=5600202211188051408247, latest=0x9d9f…34a5, have_latest=false, state=Idle
1752:2018-07-26 15:51:39  IO Worker #3 TRACE sync  peer 0 is not suitable for asking old blocks
1753-2018-07-26 15:51:39  IO Worker #3 TRACE sync  Deactivating peer 0

Note that 0x12f966affeebce9f577 == 5600202211188051408247

We would expect to be able to download old blocks from this peer. This also has the effect of not downloading old blocks from peers even if they are only a few behind.

@5chdn 5chdn added this to the 2.1 milestone Sep 11, 2018
@5chdn 5chdn added F2-bug 🐞 The client fails to follow expected behavior. P5-sometimesoon 🌲 Issue is worth doing soon. M4-core ⛓ Core client code / Rust. labels Sep 12, 2018
ascjones added a commit that referenced this issue Sep 17, 2018
#9226 might be too
permissive and causing the behaviour of the retraction soon after the
fork block. With this change the peer difficulty has to be greater than
or euqal to our syncing difficulty, so should still fix
#9225
5chdn pushed a commit that referenced this issue Oct 9, 2018
* Log block set in block_sync for easier debugging

* logging macros

* Match no args in sync logging macros

* Add QueueFull error

* Only allow importing headers if the first matches requested

* WIP

* Test for chain head gaps and log

* Calc distance even with 2 heads

* Revert previous commits, preparing simple fix

This reverts commit 5f38aa8.

* Reject headers with no gaps when ChainHead

* Reset block sync download when queue full

* Simplify check for subchain heads

* Add comment to explain subchain heads filter

* Fix is_subchain_heads check and comment

* Prevent premature round completion after restart

This is a problem on mainnet where multiple stale peer requests will
force many rounds to complete quickly, forcing the retraction.

* Reset stale old blocks request after queue full

* Revert "Reject headers with no gaps when ChainHead"

This reverts commit 0eb8655.

* Add BlockSet to BlockDownloader logging

Currently it is difficult to debug this because there are two instances,
one for OldBlocks and one for NewBlocks. This adds the BlockSet to all
log messages for easy log filtering.

* Reset OldBlocks download from last enqueued

Previously when the ancient block queue was full it would restart the
download from the last imported block, so the ones still in the queue would be
redownloaded. Keeping the existing downloader instance and just
resetting it will start again from the last enqueued block.:wq

* Ignore expired Body and Receipt requests

* Log when ancient block download being restarted

* Only request old blocks from peers with >= difficulty

#9226 might be too
permissive and causing the behaviour of the retraction soon after the
fork block. With this change the peer difficulty has to be greater than
or euqal to our syncing difficulty, so should still fix
#9225

* Some logging and clear stalled blocks head

* Revert "Some logging and clear stalled blocks head"

This reverts commit 757641d.

* Reset stalled header if useless more than once

* Store useless headers in HashSet

* Add sync target to logging macro

* Don't disable useless peer and fix log macro

* Clear useless headers on reset and comments

* Use custom error for collecting blocks

Previously we resued BlockImportError, however only the Invalid case and
this made little sense with the QueueFull error.

* Remove blank line

* Test for reset sync after consecutive useless headers

* Don't reset after consecutive headers when chain head

* Delete commented out imports

* Return DownloadAction from collect_blocks instead of error

* Don't reset after round complete, was causing test hangs

* Add comment explaining reset after useless

* Replace HashSet with counter for useless headers

* Refactor sync reset on bad block/queue full

* Add missing target for log message

* Fix compiler errors and test after merge

* ethcore: revert ethereum tests submodule update
andresilva pushed a commit that referenced this issue Oct 10, 2018
* Log block set in block_sync for easier debugging

* logging macros

* Match no args in sync logging macros

* Add QueueFull error

* Only allow importing headers if the first matches requested

* WIP

* Test for chain head gaps and log

* Calc distance even with 2 heads

* Revert previous commits, preparing simple fix

This reverts commit 5f38aa8.

* Reject headers with no gaps when ChainHead

* Reset block sync download when queue full

* Simplify check for subchain heads

* Add comment to explain subchain heads filter

* Fix is_subchain_heads check and comment

* Prevent premature round completion after restart

This is a problem on mainnet where multiple stale peer requests will
force many rounds to complete quickly, forcing the retraction.

* Reset stale old blocks request after queue full

* Revert "Reject headers with no gaps when ChainHead"

This reverts commit 0eb8655.

* Add BlockSet to BlockDownloader logging

Currently it is difficult to debug this because there are two instances,
one for OldBlocks and one for NewBlocks. This adds the BlockSet to all
log messages for easy log filtering.

* Reset OldBlocks download from last enqueued

Previously when the ancient block queue was full it would restart the
download from the last imported block, so the ones still in the queue would be
redownloaded. Keeping the existing downloader instance and just
resetting it will start again from the last enqueued block.:wq

* Ignore expired Body and Receipt requests

* Log when ancient block download being restarted

* Only request old blocks from peers with >= difficulty

#9226 might be too
permissive and causing the behaviour of the retraction soon after the
fork block. With this change the peer difficulty has to be greater than
or euqal to our syncing difficulty, so should still fix
#9225

* Some logging and clear stalled blocks head

* Revert "Some logging and clear stalled blocks head"

This reverts commit 757641d.

* Reset stalled header if useless more than once

* Store useless headers in HashSet

* Add sync target to logging macro

* Don't disable useless peer and fix log macro

* Clear useless headers on reset and comments

* Use custom error for collecting blocks

Previously we resued BlockImportError, however only the Invalid case and
this made little sense with the QueueFull error.

* Remove blank line

* Test for reset sync after consecutive useless headers

* Don't reset after consecutive headers when chain head

* Delete commented out imports

* Return DownloadAction from collect_blocks instead of error

* Don't reset after round complete, was causing test hangs

* Add comment explaining reset after useless

* Replace HashSet with counter for useless headers

* Refactor sync reset on bad block/queue full

* Add missing target for log message

* Fix compiler errors and test after merge

* ethcore: revert ethereum tests submodule update
andresilva pushed a commit that referenced this issue Oct 10, 2018
* Log block set in block_sync for easier debugging

* logging macros

* Match no args in sync logging macros

* Add QueueFull error

* Only allow importing headers if the first matches requested

* WIP

* Test for chain head gaps and log

* Calc distance even with 2 heads

* Revert previous commits, preparing simple fix

This reverts commit 5f38aa8.

* Reject headers with no gaps when ChainHead

* Reset block sync download when queue full

* Simplify check for subchain heads

* Add comment to explain subchain heads filter

* Fix is_subchain_heads check and comment

* Prevent premature round completion after restart

This is a problem on mainnet where multiple stale peer requests will
force many rounds to complete quickly, forcing the retraction.

* Reset stale old blocks request after queue full

* Revert "Reject headers with no gaps when ChainHead"

This reverts commit 0eb8655.

* Add BlockSet to BlockDownloader logging

Currently it is difficult to debug this because there are two instances,
one for OldBlocks and one for NewBlocks. This adds the BlockSet to all
log messages for easy log filtering.

* Reset OldBlocks download from last enqueued

Previously when the ancient block queue was full it would restart the
download from the last imported block, so the ones still in the queue would be
redownloaded. Keeping the existing downloader instance and just
resetting it will start again from the last enqueued block.:wq

* Ignore expired Body and Receipt requests

* Log when ancient block download being restarted

* Only request old blocks from peers with >= difficulty

#9226 might be too
permissive and causing the behaviour of the retraction soon after the
fork block. With this change the peer difficulty has to be greater than
or euqal to our syncing difficulty, so should still fix
#9225

* Some logging and clear stalled blocks head

* Revert "Some logging and clear stalled blocks head"

This reverts commit 757641d.

* Reset stalled header if useless more than once

* Store useless headers in HashSet

* Add sync target to logging macro

* Don't disable useless peer and fix log macro

* Clear useless headers on reset and comments

* Use custom error for collecting blocks

Previously we resued BlockImportError, however only the Invalid case and
this made little sense with the QueueFull error.

* Remove blank line

* Test for reset sync after consecutive useless headers

* Don't reset after consecutive headers when chain head

* Delete commented out imports

* Return DownloadAction from collect_blocks instead of error

* Don't reset after round complete, was causing test hangs

* Add comment explaining reset after useless

* Replace HashSet with counter for useless headers

* Refactor sync reset on bad block/queue full

* Add missing target for log message

* Fix compiler errors and test after merge

* ethcore: revert ethereum tests submodule update
5chdn pushed a commit that referenced this issue Oct 10, 2018
* produce portable binaries (#9725)

* HF in POA Core (2018-10-22) (#9724)

poanetwork/poa-chain-spec#87

* Use static call and apparent value transfer for block reward contract code (#9603)

* Verify block syncing responses against requests (#9670)

* sync: Validate received BlockHeaders packets against stored request.

* sync: Validate received BlockBodies and BlockReceipts.

* sync: Fix broken tests.

* sync: Unit tests for BlockDownloader::import_headers.

* sync: Unit tests for import_{bodies,receipts}.

* tests: Add missing method doc.

* Fix ancient blocks sync (#9531)

* Log block set in block_sync for easier debugging

* logging macros

* Match no args in sync logging macros

* Add QueueFull error

* Only allow importing headers if the first matches requested

* WIP

* Test for chain head gaps and log

* Calc distance even with 2 heads

* Revert previous commits, preparing simple fix

This reverts commit 5f38aa8.

* Reject headers with no gaps when ChainHead

* Reset block sync download when queue full

* Simplify check for subchain heads

* Add comment to explain subchain heads filter

* Fix is_subchain_heads check and comment

* Prevent premature round completion after restart

This is a problem on mainnet where multiple stale peer requests will
force many rounds to complete quickly, forcing the retraction.

* Reset stale old blocks request after queue full

* Revert "Reject headers with no gaps when ChainHead"

This reverts commit 0eb8655.

* Add BlockSet to BlockDownloader logging

Currently it is difficult to debug this because there are two instances,
one for OldBlocks and one for NewBlocks. This adds the BlockSet to all
log messages for easy log filtering.

* Reset OldBlocks download from last enqueued

Previously when the ancient block queue was full it would restart the
download from the last imported block, so the ones still in the queue would be
redownloaded. Keeping the existing downloader instance and just
resetting it will start again from the last enqueued block.:wq

* Ignore expired Body and Receipt requests

* Log when ancient block download being restarted

* Only request old blocks from peers with >= difficulty

#9226 might be too
permissive and causing the behaviour of the retraction soon after the
fork block. With this change the peer difficulty has to be greater than
or euqal to our syncing difficulty, so should still fix
#9225

* Some logging and clear stalled blocks head

* Revert "Some logging and clear stalled blocks head"

This reverts commit 757641d.

* Reset stalled header if useless more than once

* Store useless headers in HashSet

* Add sync target to logging macro

* Don't disable useless peer and fix log macro

* Clear useless headers on reset and comments

* Use custom error for collecting blocks

Previously we resued BlockImportError, however only the Invalid case and
this made little sense with the QueueFull error.

* Remove blank line

* Test for reset sync after consecutive useless headers

* Don't reset after consecutive headers when chain head

* Delete commented out imports

* Return DownloadAction from collect_blocks instead of error

* Don't reset after round complete, was causing test hangs

* Add comment explaining reset after useless

* Replace HashSet with counter for useless headers

* Refactor sync reset on bad block/queue full

* Add missing target for log message

* Fix compiler errors and test after merge

* ethcore: revert ethereum tests submodule update

* Add hardcoded headers (#9730)

* add foundation hardcoded header #6486017

* add ropsten hardcoded headers #4202497

* add kovan hardcoded headers #9023489

* gitlab ci: releasable_branches: change variables condition to schedule (#9729)
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
F2-bug 🐞 The client fails to follow expected behavior. M4-core ⛓ Core client code / Rust. P5-sometimesoon 🌲 Issue is worth doing soon.
Projects
None yet
Development

No branches or pull requests

2 participants