SQLite is a foundation of offline, local-first and edge deployed software. Wouldn't it be great, however, if we could merge two or more SQLite databases together and not run into any conflicts?

This project implements CRDTs and CRRs in SQLite, allowing databases that share a common schema to merge their state together. Merges can happen between an arbitrary number of peers and all peers will eventually converge to the same state.

cfsqlite works by adding metadata tables and triggers around your existing database schema. This means that you do not have to change your schema in order to get conflict resolution support -- with a few caveats around uniqueness constraints and foreign keys. See Schema Design for CRDTs & Eventual Consistency.

I'm working on a demo application. You can, of course, check out the repo and repeat what occurs in this video to play with this locally.

git clone conflict-free-sqlite
cd conflict-free-sqlite/prototype
pnpm install
cd demo-env
pnpm build
node

You can view a conflict-free DB in action in the __tests__ folder of the replicator package: https://github.com/tantaman/conflict-free-sqlite/blob/main/prototype/replicator/src/__tests__/merge-random-2.test.ts

Auto-migration of an existing sqlite db to be conflict free is not yet implemented. This will live in the migrator package. A manual migration of a database of todos can be seen here: https://github.com/tantaman/conflict-free-sqlite/tree/main/prototype/test-schemas

In the future the steps to using cfsqlite will be:

1. Run migrator to migrate an existing DB to a conflict-free schema
2. Pull in the replicator API for your target language
3. Connect the replicator to peer databases

https://munin.uit.no/bitstream/handle/10037/22344/thesis.pdf?sequence=2

cfsqlite improves upon [1] in the following ways --

• [1] stores two copies of all the data. cfsqlite only keeps one by leveraging views and ISNTEAD OF triggers.
• [1] cannot compute deltas between databases without sending the full copy of each database to be compared. cfsqlite only needs the logical clock (1 64bit int per peer) of a given database to determine what updates that database is missing.

https://hal.inria.fr/hal-02983557/document

cfsqlite improves upon [2] in the following ways --

• [2] is implemented in a specific ORM. cfsqlite runs at the db layer and allows existing applications to interface with the db as normal.
• [2] keeps a queue of all writes. This queue is drained when those writes are merged. This means that [2] can only sync changes to a single centralized node. cfsqlite keeps a logical clock at each database. If a new database comes online it sends its logical clock to a peer. That peer can compute what changes are missing from the clock.

https://www.youtube.com/watch?v=DEcwa68f-jY

cfsqlite improves upon [3] in the following ways --

• [3] isn't really relational it all. It saves all data in a single table and is using sqlite as a key-value store. As such, it cannot work with your existing database schema. cfsqlite builds around your existing schemas.

These projects helped improve my understanding of CRDTs on this journey --

• shelf
• tiny-merge
• Merkle-CRDT
  • Not used yet but might become an alternative clock implementation for use cases with unbounded numbers of peers

cfsqlite currently does not support:

1. Foreign key cosntraints. You can still have foreign keys (i.e. a column with an id of another row), but they can't be enforced by the db.
   1. TODO: discuss design alternatives and why this is actually not a bad thing when considering row level security.
2. Uniqueness constraints other than the primary key. The only enforceably unique column in a table should be the primary key. Other columns may be indices but they may not be unique.
   1. TODO: discuss this in much more detail.

Note: prior art [1] & [2] claim to support foreign key and uniqueness constraints. I believe their approach is unsound and results in update loops and have not incoroprated it into cfsqlite. If I'm wrong, I'll gladly fold their approach in.

Tables are modeled as GSets where each item has a causal length. You can call this a "CLSet". This allows us to keep all rows as well as track deletes so your application will not see deleted rows.

Rows are currently modeled as LWWW maps. I.e., each column in a row is a LWW Register.

Things to support in the future

• counter columns
• MVR (multi-value register) columns

Deltas between databases are calculated by each database keeping a version vector.

Every row in the database is associated with a copy of the version vector. This copy is a snapshot of the value of the vector at the time the most recent write was made to the row.

If DB-A wants changes from DB-B,

• DB-A sends its version vector to DB-B
• DB-B finds all rows for which any element in the row's snapshot vector is greater than the corresponding element in the provided vector or for which the provided vector is missing an entry (https://github.com/tantaman/conflict-free-sqlite/blob/main/prototype/replicator/src/queries.ts#L59-L63)
• DB-B sends these rows to DB-A
• DB-A applys the changes
• DB-A now has all of DB-B's updates

This algorithm requires causal delivery of message during the time which two peers decide to sync.

cfsqlite is currently implemented as a set of views, triggers, and conflict free base tables.

The views match an application's existing database schema so little to no changes need be made to existing applications.

Whenever sqlite tries to write to a view, we intercept that write and write it to the conflict free base tables instead. This allows you to issue arbitrarily complex writes (e.g. UPDATE x WHERE condition) as SQLite will resolve the impacted rows via its query engine.

You can view a set of manually constructed view and triggers here: https://github.com/tantaman/conflict-free-sqlite/tree/main/prototype/test-schemas

cfsqlite is currently 2-3x slower than base sqlite. I believe we can get perf to be near identical. The current bottlenecks are:

1. The current database clock value is stored in a table and must be touched every write
2. The site id of the database is stored in a table and queried every write

We can move both of these values out of their tables and into a variable in-memory. Preliminary tests show that doing this results in near identical perf to sqlite.

• Sharing & Privacy -- in a real-world collaborative scenario, you may not want to share your entire database with other peers. Thus, in addition to clock information, we must keep visibility information to use when computing deltas and doing replication.
• Byzantine fault tolerance -- cfsqlite currently assumes friendly actors. We need to guard against malicious updates.
• Subselects -- peers may want to sync subsets of the database even if they have access to the entire thing. Compute deltas but only send those deltas that fall into the peer's provided query.

Say we have a databse schema called "Animal App." Alice, Bob and Billy all have local copies of "Animal App" on their devices. They start their day at a hostel with all of their devices synced. They then part ways, backpacking into the wilderness each with their own copy of the db.

As they see different (or maybe even the same) animals, they record their observations.

• Name
• Genus
• Species
• Habitat
• Diet
• Etc.

Some observations may even involve making updates to the prior day's (or week's) observations written by other members of the party.

At the end of the day, the group comes back together. They need to merge all of their work. cfsqlite will allow Alice, Bob and Billy to merge their changes (without conflict) in a p2p fashion and converge to the same state.

Note that "without conflict" would be based on the rules of the selected CRDTs used within the schema.

Some example are --

• Tables might be grow only sets -- thus never losing an observation.
  • Or sets with a causal length so we can safely remove rows
• Table columns might be last write win (LWW) registers -- converging but throwing out earlier writes
• Table columns might be multi value (MV) registers -- keeping around all concurrent edits to a single column for users (or code) to pick and merge later.
• A column might be a counter CRDT which accumulates all observations from all parties