If we are low on addresses for peering, we need to discover more peers. The
previous behavior would query existing peers; however, if an existing peer
does not participate in peer exchange, then our node will not discover more peers.
This change consults both existing peers as well as seeds when there is a deficit
in address book addresses. This allows for discovering peers though existing channels
as well as via seeds if existing peers do not share addresses.
The node.NewNode method is pretty complex at the moment, an in order to address issues like #3156, we need to simplify the interface for partial node instantiation. In some places, we don't need to build up a full node (like in the node.TestCreateProposalBlock test), but the complexity of such partial instantiation needs to be reduced.
This PR aims to eventually make this easier/simpler.
See also this gist https://gist.github.com/thanethomson/56e1640d057a26186e38ad678a1d114c for some background work done when starting to refactor here.
## Commits:
* [WIP] Refactor node.NewNode to simplify
The `node.NewNode` method is pretty complex at the moment, an in order
to address issues like #3156, we need to simplify the interface for
partial node instantiation. In some places, we don't need to build up a
full node (like in the `node.TestCreateProposalBlock` test), but the
complexity of such partial instantiation needs to be reduced.
This PR aims to eventually make this easier/simpler.
* Refactor state loading and genesis doc provider into state package
* Refactor for clarity of return parameters
* Fix incorrect capitalization of error messages
* Simplify extracted functions' names
* Document optionally-prefixed functions
* Refactor optionallyFastSync for clarity of separation of concerns
* Restructure function for early return
* Restructure function for early return
* Remove dependence on deprecated panic functions
* refactor code a bit more
plus, expose PEXReactor on node
* align logger names
* add a changelog entry
* align logger names 2
* add a note about PEXReactor returning nil
* Remove deprecated PanicXXX functions from codebase
As per discussion over
[here](https://github.com/tendermint/tendermint/pull/3456#discussion_r278423492),
we need to remove these `PanicXXX` functions and eliminate our
dependence on them. In this PR, each and every `PanicXXX` function call
is replaced with a simple `panic` call.
* add a changelog entry
Should fix#3451, #2723 and #3317.
Test TestResetTimeoutPrecommitUponNewHeight is simplified so it reduces a risk of timeout failure. Furthermore, timeout we wait for TimeoutEvents is increased, and the timeout value is more precisely computed. This should hopefully decrease a chance of non-deterministic test failures.
This assertion is problematic to ensure consistently due to dependency on scheduler. On the other hand, if I am not wrong, order in which messages are read from the channel respects order in which messages are written. Therefore, process will receive 2f+1 precommits that are not all for v (one is for nil) so TriggeredTimeoutPrecommit would be set to true. So we don't need to assert it. I know that it would be better to still assert to it but I don't know how to do it without sleep and that is ugly and is causing us nondeterministic failures.
* check every block appHash
Fixes#3460
Not really needed, but it would detect if the app changed in a way it
shouldn't have.
* add a changelog entry
* no need to return an error if we panic
* rename methods
* rename methods once again
* add a test
* correct error msg
plus fix a few go-lint warnings
* better panic msg
This should fix#3576 (ran it many times locally but only time will tell). The test actually only checked for the opcode of the error. From the name of the test we actually want to test if we see a timeout after a pre-defined time.
## Commits:
* increase readWrite timeout as it is also used in the `Accept` of the tcp
listener:
- before this caused the readWriteTimeout to kick in (rarely) while Accept
- as a side-effect: remove obsolete time.Sleep: in both listener cases
the Accept will only return after successfully accepting and the timeout
that is supposed to be tested here will be triggered because there is a
read without a write
- see if we actually run into a timeout error (the whole purpose of
this test AFAIU)
Signed-off-by: Ismail Khoffi <Ismail.Khoffi@gmail.com>
* makee local test-vars `const`
Signed-off-by: Ismail Khoffi <Ismail.Khoffi@gmail.com>
## Additional comments:
@xla: Confusing how an accept could take longer than that, but assuming a noisy environment full of little docker whales will be slower than what 50 years of silicon are capable of. The only thing I'd be vary of is that we mask structural issues with the code by just bumping the timeout, if we are sensitive towards that it warrants invesigation, but again this might only be true in the environment our CI runs in.
(#2611) had suggested that an iterative version of
SimpleHashFromByteSlice would be faster, presumably because
we can envision some overhead accumulating from stack
frames and function calls. Additionally, a recursive algorithm risks
hitting the stack limit and causing a stack overflow should the tree
be too large.
Provided here is an iterative alternative, a simple test to assert
correctness and a benchmark. On the performance side, there appears to
be no overall difference:
```
BenchmarkSimpleHashAlternatives/recursive-4 20000 77677 ns/op
BenchmarkSimpleHashAlternatives/iterative-4 20000 76802 ns/op
```
On the surface it might seem that the additional overhead is due to
the different allocation patterns of the implementations. The recursive
version uses a single `[][]byte` slices which it then re-slices at each level of the tree.
The iterative version reproduces `[][]byte` once within the function and
then rewrites sub-slices of that array at each level of the tree.
Eexperimenting by modifying the code to simply calculate the
hash and not store the result show little to no difference in performance.
These preliminary results suggest:
1. The performance of the current implementation is pretty good
2. Go has low overhead for recursive functions
3. The performance of the SimpleHashFromByteSlice routine is dominated
by the actual hashing of data
Although this work is in no way exhaustive, point #3 suggests that
optimizations of this routine would need to take an alternative
approach to make significant improvements on the current performance.
Finally, considering that the recursive implementation is easier to
read, it might not be worthwhile to switch to a less intuitive
implementation for so little benefit.
* re-add slice re-writing
* [crypto] Document SimpleHashFromByteSlicesIterative
Continues from #3280 in building support for batched requests/responses in the JSON RPC (as per issue #3213).
* Add JSON RPC batching for client and server
As per #3213, this adds support for [JSON RPC batch requests and
responses](https://www.jsonrpc.org/specification#batch).
* Add additional checks to ensure client responses are the same as results
* Fix case where a notification is sent and no response is expected
* Add test to check that JSON RPC notifications in a batch are left out in responses
* Update CHANGELOG_PENDING.md
* Update PR number now that PR has been created
* Make errors start with lowercase letter
* Refactor batch functionality to be standalone
This refactors the batching functionality to rather act in a standalone
way. In light of supporting concurrent goroutines making use of the same
client, it would make sense to have batching functionality where one
could create a batch of requests per goroutine and send that batch
without interfering with a batch from another goroutine.
* Add examples for simple and batch HTTP client usage
* Check errors from writer and remove nolinter directives
* Make error strings start with lowercase letter
* Refactor examples to make them testable
* Use safer deferred shutdown for example Tendermint test node
* Recompose rpcClient interface from pre-existing interface components
* Rename WaitGroup for brevity
* Replace empty ID string with request ID
* Remove extraneous test case
* Convert first letter of errors.Wrap() messages to lowercase
* Remove extraneous function parameter
* Make variable declaration terse
* Reorder WaitGroup.Done call to help prevent race conditions in the face of failure
* Swap mutex to value representation and remove initialization
* Restore empty JSONRPC string ID in response to prevent nil
* Make JSONRPCBufferedRequest private
* Revert PR hard link in CHANGELOG_PENDING
* Add client ID for JSONRPCClient
This adds code to automatically generate a randomized client ID for the
JSONRPCClient, and adds a check of the IDs in the responses (if one was
set in the requests).
* Extract response ID validation into separate function
* Remove extraneous comments
* Reorder fields to indicate clearly which are protected by the mutex
* Refactor for loop to remove indexing
* Restructure and combine loop
* Flatten conditional block for better readability
* Make multi-variable declaration slightly more readable
* Change for loop style
* Compress error check statements
* Make function description more generic to show that we support different protocols
* Preallocate memory for request and result objects
* use dialPeer function in a seed mode
Fixes#3532
by storing a number of attempts we've tried to connect in-memory and
removing the address from addrbook when number of attempts > 16
Fixes#3457
The topic of the issue is that : write a BlockRequest int requestsCh channel will create an timer at the same time that stop the peer 15s later if no block have been received . But pop a BlockRequest from requestsCh and send it out may delay more than 15s later. So that the peer will be stopped for error("send nothing to us").
Extracting requestsCh into its own goroutine can make sure that every BlockRequest been handled timely.
Instead of the requestsCh handling, we should probably pull the didProcessCh handling in a separate go routine since this is the one "starving" the other channel handlers. I believe the way it is right now, we still have issues with high delays in errorsCh handling that might cause sending requests to invalid/ disconnected peers.
What happened:
New code was supposed to fall back to last height changed when/if it
failed to find validators at checkpoint height (to make release
non-breaking).
But because we did not check if validator set is empty, the fall back
logic was never executed => resulting in LoadValidators returning an
empty validator set for cases where `lastStoredHeight` is checkpoint
height (i.e. almost all heights if the application does not change
validator set often).
How it was found:
one of our users - @sunboshan reported a bug here
https://github.com/tendermint/tendermint/pull/3537#issuecomment-482711833
* use last height changed in validator set is empty
* add a changelog entry
A prior change to address accidental DNS lookups introduced the
SocketAddr on peer, which was then used to add it to the addressbook.
Which in turn swallowed the self reported port of the peer, which is
important on a reconnect. This change revives the NetAddress on NodeInfo
which the Peer carries, but now returns an error to avoid nil
dereferencing another issue observed in the past. Additionally we could
potentially address #3532, yet the original problem statemenf of that
issue stands.
As a drive-by optimisation `MarkAsGood` now takes only a `p2p.ID` which
makes it interface a bit stricter and leaner.
* rpc: store validator info periodly
* increase ValidatorSetStoreInterval
also
- unexpose it
- add a comment
- refactor code
- add a benchmark, which shows that 100000 results in ~ 100ms to get 100
validators
* make the change non-breaking
* expand comment
* rename valSetStoreInterval to valSetCheckpointInterval
* change the panic msg
* add a test and changelog entry
* update changelog entry
* update changelog entry
* add a link to PR
* fix test
* Update CHANGELOG_PENDING.md
Co-Authored-By: melekes <anton.kalyaev@gmail.com>
* update comment
* use MaxInt64 func
* add actionable advice for ErrAddrBookNonRoutable err
Should replace https://github.com/tendermint/tendermint/pull/3463
* reorder checks in addrbook#addAddress so
ErrAddrBookPrivate is returned first
and do not log error in DialPeersAsync if the address is private
because it's not an error
ListOfKnownAddresses is removed
panic if addrbook size is less than zero
CrawlPeers does not attempt to connect to existing or peers we're currently dialing
various perf. fixes
improved tests (though not complete)
move IsDialingOrExistingAddress check into DialPeerWithAddress (Fixes#2716)
* addrbook: preallocate memory when saving addrbook to file
* addrbook: remove oldestFirst struct and check for ID
* oldestFirst replaced with sort.Slice
* ID is now mandatory, so no need to check
* addrbook: remove ListOfKnownAddresses
GetSelection is used instead in seed mode.
* addrbook: panic if size is less than 0
* rewrite addrbook#saveToFile to not use a counter
* test AttemptDisconnects func
* move IsDialingOrExistingAddress check into DialPeerWithAddress
* save and cleanup crawl peer data
* get rid of DefaultSeedDisconnectWaitPeriod
* make linter happy
* fix TestPEXReactorSeedMode
* fix comment
* add a changelog entry
* Apply suggestions from code review
Co-Authored-By: melekes <anton.kalyaev@gmail.com>
* rename ErrDialingOrExistingAddress to ErrCurrentlyDialingOrExistingAddress
* lowercase errors
* do not persist seed data
pros:
- no extra files
- less IO
cons:
- if the node crashes, seed might crawl a peer too soon
* fixes after Ethan's review
* add a changelog entry
* we should only consult Switch about peers
checking addrbook size does not make sense since only PEX reactor uses
it for dialing peers!
https://github.com/tendermint/tendermint/pull/3011#discussion_r270948875
* OriginalAddr -> SocketAddr
OriginalAddr records the originally dialed address for outbound peers,
rather than the peer's self reported address. For inbound peers, it was
nil. Here, we rename it to SocketAddr and for inbound peers, set it to
the RemoteAddr of the connection.
* use SocketAddr
Numerous places in the code call peer.NodeInfo().NetAddress().
However, this call to NetAddress() may perform a DNS lookup if the
reported NodeInfo.ListenAddr includes a name. Failure of this lookup
returns a nil address, which can lead to panics in the code.
Instead, call peer.SocketAddr() to return the static address of the
connection.
* remove nodeInfo.NetAddress()
Expose `transport.NetAddress()`, a static result determined
when the transport is created. Removing NetAddress() from the nodeInfo
prevents accidental DNS lookups.
* fixes from review
* linter
* fixes from review
Anton Kaliaev
Ivan Kushmantsev
JamesRay
Roman Shtylman
Ethan Buchman
Thane Thomson
Zarko Milosevic
Greg Hill
Ismail Khoffi
Sean Braithwaite
kevlubkcm
hucc
zjubfd
dongsamb
Martin Dyring-Andersen
Alexander Simmerl