Design Doc 011: Auditor-Exchange Database Synchronization ######################################################### Summary ======= Ways for the auditor to obtain a current copy of the exchange database (to verify that the exchange is operating correctly) are discussed. Motivation ========== 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. Requirements ============ * 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must enable the auditor database to serve as a full backup of the exchange's database (even if possibly slightly outdated due to asynchronous replication or network outages).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. Proposed Solution ================= * 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 anyway.) * 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. Those are: * /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 exchange. * /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. A good order for replicating the tables should be: * exchange_sign_keys * signkey_revocations * auditors * denominations * denomination_revocations * auditor_denom_sigs * reserves * reserves_out * reserves_in * reserves_close * known_coins * deposits * refunds * refresh_commitments * refresh_transfer_keys * refresh_revealed_coins * recoup_refresh * recoup Alternatives ============ * 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. Drawbacks ========= * SERIAL IDs required in all tables that are "append-only" / immutable. * Additional custom logic required to recompute mutable tables on-demand. * 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 ================