Instead of searching from block 0, we now keep a record of
base trim height in the DB itself.
Also, we no longer trim the latest AT state for non-finished ATs
in case they are in deep sleeping and we need their state for when
they awaken.
Symptoms are:
* db/blockchain.log is pretty much exactly 50MB - the checkpoint-triggering size.
* Loads of threads are stuck waiting for HSQLDB's CountUpDownLatch$Sync.await()
* Synchronizer, or some other thread, possibly orphaning blocks.
The cause seems to be method A, which has a repository session,
calls EventBus.INSTANCE.notify() and one of the event listeners
then obtains their own repository session to do repository 'work'.
In the meantime, the HSQLDB log has reached 50MB, triggering auto-checkpoint.
HSQLDB attempts to CHECKPOINT, but waits for existing transactions
to complete, and also blocks starting new transactions.
Thus, one of the event listeners is blocked when they try to obtain
a new repository session, but HSQLDB never performs CHECKPOINT
because the event notifier (method A) still has an unfinished
transaction - hence deadlock.
Drop created_when column from ATStates as it never changes
and can be fetched from ATs table.
This takes about 50s on a fast machine.
Correspondingly rebuild height-based index on ATStates.
This takes about 3 minutes on a fast machine.
Modify AT-related repository methods and callers.
Aggressively remove 'old' (> 2 weeks) actual AT
state binary data, leaving only the hash in DB
(for syncing purposes). Seems to keep up with
syncing from another node on localhost.
There's still an existing issue where log entries like this appear:
Unable to trim old online accounts signatures in repository
which is actually caused by:
integrity constraint violation: unique constraint or index violation; SYS_PK_10092 table: BLOCKS
which seems to be a bug in the version of HSQLDB we use.
(Tested using synced-from-scratch DB).
It's not clear what the actual problem is at this point.
It might be possible to switch to v2.5.1 if our recent HSQLDB-related
commits have fixed/worked-around the OOM issues.
Move the inner method from BlockChain to Controller.
Remove blockchain lock as it's not needed because it's not an
HSQLDB "serialization failure" but constraint violation.
Trimming old online accounts signatures limited to batches of 1440
rows to reduce CPU and memory load.
Added separate method to determine status of P2SH transactions,
returning UNFUNDED, FUNDING_IN_PROGRESS, REDEEMED, etc.
Added code to trade-bot to increase robustness. Lots more
changes including unified state change/logging, checking
for existing MESSAGEs, etc.
Added missing websocket methods to silence log noise.
Trade-bot now called per block during synchronization,
instead of per batch, to pick up edge cases where some
potential trade-bot transitions were missed, resulting
in failed trades.
Corresponding changes in Controller, such as notifying
event bus of new block in same thread (thus blocking)
instead of using executor.
Added slightly more robust common block determination
to Synchronizer.
Refactored code in BTC class to use new BitcoinException
rather than simply returning null, with added sub-classes
allowing differentiation between network issues or fund
issues.
Changed BTC.buildSpend to try harder to find UXTOs to
address false "insufficient funds" issues.
Repository change to add index on MessageTransactions
for quicker look-up of trade-related messages.
Reduced reliance on bitcoinj library in BTCP2SH.
Reworked ElectrumX to better detect errors rather than
continuously try more servers to no avail.
Also added genesis block check in case of servers on
different Bitcoin networks.
Now tries to extract upstream bitcoind error codes
and pass those up to caller via exceptions.
Updated list of testnet servers.
MemoryPoW now detects thread interrupt and exits fast.
Moved some non-generic transaction-related repository
methods to their own subclass. For example:
moved TransactionRepository.getMessagesByRecipient
to MessageRepository.getMessagesByParticipants
Updated and added more tests.
Was "run.sh" but renamed to "start.sh" to better complement "stop.sh".
"run.sh" is now a symbolic link to "start.sh"
Reworked Java version check to remove dependency on "bc" tool which
seems not to be installed on some Ubuntu distributions?
Removed -XX:NativeMemoryTracking flag from JVM args.
Fixed incorrect comment regarding java.net.preferIpV4Stack.
Fixed typo in comment.
Requires node shutdown, lots of time (10s of minutes), spare storage space.
Called via: java -cp qortal.jar org.qortal.RepositoryMaintenance
Not (yet) for general consumption.
Added Qortal-side HSQLDB PreparedStatement cache, hashed
by SQL query string, to reduce re-preparing statements.
(HSQLDB actually does the work in avoiding re-preparing
by comparing its own query-to-statement cache map, but we
need to keep an 'open' statement on our side for this to
happen).
Support added for batched INSERT/UPDATE SQL statements to
update many rows in one call.
Several specific repository calls, e.g. modifyMintedBlockCount
or modifyAssetBalance, now have batch versions that allow
many rows to be updated in one call.
In Block, when distributing block rewards, although we still
build a map of balance changes to apply after all calculations,
this map is now handed off wholesale to the repository to
apply in one (or two) queries, instead of a repository call
per account. The balanceChanges map is now keyed by account
address, as opposed to actual Account.
Also in Block, we try to cache the fetched online reward-shares
(typically in Block.isValid et al) to avoid re-fetching them
later when calculating block rewards.
In addition, actually fetching online reward-shares is no longer
done index-by-index, but the whole array of indexes is passed
wholesale to the repository which then returns the corresponding
reward-shares as a list.
In Block.increaseAccountLevels, blocks minted counts are also
updated in one single repository call, rather than one
repository call per account.
When distributing Block rewards to legacy QORA holders,
all necessary info is fetched from the repository in one hit
instead of two-phases of: 1. fetching eligible QORA holders,
and 2. fetching extra data for that QORA holder as needed.
In addition, updated QORT_FROM_QORA asset balances are done
via one batch repository call, rather than per update.
Symptoms include this in logs:
Unexpected zero effective minter level for reward-share %s - using 1 instead!
This occurs when Synchronizer compares two sub-chains from a common block,
and one of the blocks is signed by a reward-share key that has
subsequently been cancelled.
Although this is catered for, excessive log-spam is emited.
So in addition to demoting the log level from WARN to DEBUG,
more code has been added to try harder to find the actual data needed,
thus preventing the logging in the first place.
New repository transaction search method added to support above,
along with corresponding tests.
catbref
CalDescent