Design Doc 011: Auditor-Exchange Database Synchronization
Ways for the auditor to obtain a current copy of the exchange database (to
verify that the exchange is operating correctly) are discussed.
The Taler auditor is expected to check that the exchange is operating
correctly. For this, it needs (read-only) access to the exchange database and
to the bank account of the exchange. Bank account access is a matter of
setting up an additional user with limited rights with the bank and is out of
the scope of this document. For database access, the auditor should not trust
the exchange. In particular, the auditor must assume that the exchange may
violate basic database constraints (like foreign keys) or delete/alter records
maliciously. However, we also do not want to complicate the auditor logic to
cope with with violations of well-formedness constraints (like foreign keys or
non-NULL values or size constraints on fields). Finally, the mechanism by
which the auditor obtains the database must provide a reasonably current
database and the process must perform reasonably well.
* The solution must allow data to be copied incrementally.
* The solution must tolerate network outages and recover after connectivity
between exchange and auditor is restored.
* The solution must enable the auditor database to serve as a full backup
of the exchange's database (even if possibly slightly outdated due to
asynchronous replication or network outages).
* The solution must scale, in particular if the exchange shards the database,
the auditor must be also able to use the same kind of sharding and the
synchronization should be possible per shard.
* The synchronization mechanism must not allow an attacker controlling the
exchange database to delete or modify arbitrary data from the auditor's copy
via the synchronization mechanism (in other words, some tables are
append-only and unalterable).
* The solution must support database schema updates. Those may require some
downtime and closely coordinated work between exchange and auditor.
* The solution must enable eventual garbage collection at the exchange to
be permitted and replicated at the auditor (e.g. DELETE on usually append-only
tables due to a CASCADE from expired denomination keys).
* The synchronization mechanism should raise an alert if the exchange violates basic
constraints (unexpected schema changes, deletion/motification on append-only
tables) and then NOT replicate those changes. The auditor's internal asynchronous
helper may then soft-fail (log and exit) until the exchange has rectified the
problem (by manual, human intervention resulting in an exchange master database
that again maintains the required invariants).
After the corrected master database has been again synchronized with the
primary copy of the auditor, the auditor's helper is resumed and can continue to
copy the (now valid) database records into the auditor's internal version.
* A good solution would work independently of the specific database used.
* Use "common" incremental database replication (whichever is
approproate for the exchange database setup, synchronous
or asynchronous) to make a 1:1 copy of the exchange database
at the auditor. This should work for any full-featured
modern database. This "ingress" copy cannot be trusted, as constraint
violations or deletions would also be replicated.
* Use helper process to SELECT against the local "ingress" copy (by
SERIAL ID => make sure all append-only tables have one!)
to copy append-only tables to a second, "trusted" and fully
auditor-controlled copy of the database.
Order (or transactionally group) SELECT
statements to ensure foreign key constraints are maintained.
For mutable tables (basically, only current reserve balance)
do not make another copy, but do have logic to recompute mutable
tables from other data *if* we need to recover from backup.
* On schema migration, halt exchange, once auditor DB has
synchronized, update all DB schema (the "ingress" DB schema
may be updated automatically when the exchange DB schema is
migrated, but the "trusted" DB of the auditor must most likely
be manually migrated), then finally resume "ingress" to "trusted"
helper-based DB synchronization and restart the exchange.
* For GC, simply run GC logic also on auditor's "trusted" copy.
(The synchronization mechanism will take care of the primary copy,
and the helper to copy should not be disturbed by the DELETE operations
* The auditor's "ingress" database should be well isolated from
the rest of the auditor's system and database
(different user accounts). The reason is that we should not
assume that the Postgres replication code is battle-tested with
malicious parties in mind.
* The canonical Postgres synchronization between exchange and the
auditor's "ingress" database must use transport security.
The above solution does not gracefully handle mutable tables on which
the exchange performs UPDATE statements, as such updates will not bump
the BIGSERIAL and thus would not be replicated by the helper. Thus, we
need to consider all tables that the exchange ever performs UPDATE on.
* /reserves/ --- the exchange updates the remaining reserve balance;
here the auditor currently performs a sanity check against
its own reserve balance calculation. The proposed way to address
this is to make this sanity check optional and to be only used if
the auditor auditor runs against the "primary" exchange database
(like an internal audit). This is acceptable, as an inaccurate
reserve balance is mostly used to raise an early warning and not
indicative of any actualized financial gains or losses from the
* /deposits/ --- the exchange updates the /tiny/ and /done/ bit
fields. /tiny/ can be trivially established by the auditor, and
we can simply avoid the auditor considering that bit. /done/
was so far only used to enrich the reporting. The proposed way
to address the uses of both fields is thus to only use them in
internal audits (against the primary exchange database). Both
can be safely ignored by the external audit.
* /prewire/ --- the exchange updates the /finished/ and /failed/
bits. The entire table is not used by the auditor and its
main values cannot be validated by the auditor anyway.
* /auditors/ --- the exchange updates the /is_active/ and /last_change/
fields. The entire table is of no concern to the auditor.
* Copy the Postgres WAL, filter it for "illegal" operations
and then apply it at the auditor end. Disadvantages: WAL
filtering is not a common operation (format documented?),
this would be highly Postgres-specific, and would require
complex work to write the filter. Also unsure how one
could later recover gracefully from transient errors
(say where the exchange recified a bogus DELETE).
* Directly SELECT against the (remote) exchange DB and then
INSERT/UPDATE at the auditor's local copy. Disadvantages:
remote SELECT likely very expensive due to high latency.
Diagnostics more difficult. May expose exchange to additional
risks from auditor, such as attacks exhausting DB resources
by running expensive SELECTs.
* SERIAL IDs required in all tables that are "append-only" / immutable.
* Additional custom logic required to recompute mutable tables
* Limited ability to cope with mutable tables, imposes restrictions
on future exchange database evolution.
* Helper logic to SELECT data in batches that will certainly
maintain invariants may be a bit tricky, but in principle
the foreign key constraints should form a DAG, simply dictating
the order in which new entries are to be copied. It may also
be that simply running "big" transactions across all tables
is the answer, to be investigated what performs better.
* A malicious exchange could theoretically send expensive transactions
to the auditor via the replication mechanism (possibly ones that
it did not even execute locally itself) to DoS the "ingress"
database. This would be noticed primarily by load
monitoring or even the auditor lagging unusually far behind the
exchange's transaction history. We believe this is acceptable,
as it would imply highly visible malicious exchange behavior for
virtually no significant gain.
* The proposed solution does not create a transactional, synchronous
write-only log as suggested by CodeBlau (see audit report, Section 9.4).
We believe doing so would be overly costly, both in terms of
complexity and performance, for limited gains.
Discussion / Q&A