The last transaction store was a great example of how not to write
scalable software. For a variety of reasons, it was very slow at
processing transaction inserts. Among them:
1) Every single transaction record being saved in a linked list
(container/list), and inserting into this list would be an O(n)
operation so that records could be ordered by receive date.
2) Every single transaction in the above mentioned list was iterated
over in order to find double spends which must be removed. It is
silly to do this check for mined transactions, which already have
been checked for this by btcd. Worse yet, if double spends were
found, the list would be iterated a second (or third, or fourth)
time for each removed transaction.
3) All spend tracking for signed-by-wallet transactions was found on
each transaction insert, even if the now spent previous transaction
outputs were known by the caller.
This list could keep going on, but you get the idea. It was bad.
To resolve these issues a new transaction store had to be implemented.
The new implementation:
1) Tracks mined and unmined transactions in different data structures.
Mined transactions are cheap to track because the required double
spend checks have already been performed by the chain server, and
double spend checks are only required to be performed on
newly-inserted mined transactions which may conflict with previous
unmined transactions.
2) Saves mined transactions grouped by block first, and then by their
transaction index. Lookup keys for mined transactions are simply
the block height (in the best chain, that's all we save) and index
of the transaction in the block. This makes looking up any
arbitrary transaction almost an O(1) operation (almost, because
block height and block indexes are mapped to their slice indexes
with a Go map).
3) Saves records in each transaction for whether the outputs are
wallet credits (spendable by wallet) and for whether inputs debit
from previous credits. Both structures point back to the source
or spender (credits point to the transaction that spends them, or
nil for unspent credits, and debits include keys to lookup the
transaction credits they spent. While complicated to keep track
of, this greatly simplifies the spent tracking for transactions
across rollbacks and transaction removals.
4) Implements double spend checking as an almost O(1) operation. A
Go map is used to map each previous outpoint for all unconfirmed
transactions to the unconfirmed tx record itself. Checking for
double spends on confirmed transaction inserts only involves
looking up each previous outpoint of the inserted tx in this map.
If a double spend is found, removal is simplified by only
removing the transaction and its spend chain from store maps,
rather than iterating a linked list several times over to remove
each dead transaction in the spend chain.
5) Allows the caller to specify the previous credits which are spent
by a debiting transaction. When a transaction is created by
wallet, the previous outputs are already known, and by passing
their record types to the AddDebits method, lookups for each
previously unspent credit are omitted.
6) Bookkeeps all blocks with transactions with unspent credits, and
bookkeeps the transaction indexes of all transactions with unspent
outputs for a single block. For the case where the caller adding a
debit record does not know what credits a transaction debits from,
these bookkeeping structures allow the store to only consider known
unspent transactions, rather than searching through both spent and
unspents.
7) Saves amount deltas for the entire balance as a result of each
block, due to transactions within that block. This improves the
performance of calculating the full balance by not needing to
iterate over every transaction, and then every credit, to determine
if a credit is spent or unspent. When transactions are moved from
unconfirmed to a block structure, the amount deltas are incremented
by the amount of all transaction credits (both spent and unspent)
and debited by the total amount the transaction spends from
previous wallet credits. For the common case of calculating a
balance with just one confirmation, the only involves iterating
over each block structure and adding the (possibly negative)
amount delta. Coinbase rewards are saved similarly, but with a
different amount variable so they can be seperatly included or
excluded.
Due to all of the changes in how the store internally works, the
serialization format has changed. To simplify the serialization
logic, support for reading the last store file version has been
removed. Past this change, a rescan (run automatically) will be
required to rebuild the transaction history.
Josh Rickmar