|
| 1 | +# Row Level Security (RLS) with CloudSync |
| 2 | + |
| 3 | +CloudSync is fully compatible with PostgreSQL Row Level Security. Standard RLS policies work out of the box. |
| 4 | + |
| 5 | +## How It Works |
| 6 | + |
| 7 | +### Column-batch merge |
| 8 | + |
| 9 | +CloudSync resolves CRDT conflicts at the column level — a sync payload may contain individual column changes arriving one at a time. Before writing to the target table, CloudSync buffers all winning column values for the same primary key and flushes them as a single SQL statement. This ensures the database sees a complete row with all columns present. |
| 10 | + |
| 11 | +### UPDATE vs INSERT selection |
| 12 | + |
| 13 | +When flushing a batch, CloudSync chooses the statement type based on whether the row already exists locally: |
| 14 | + |
| 15 | +- **New row**: `INSERT ... ON CONFLICT DO UPDATE` — all columns are present (including the ownership column), so the INSERT `WITH CHECK` policy can evaluate correctly. |
| 16 | +- **Existing row**: `UPDATE ... SET ... WHERE pk = ...` — only the changed columns are set. The UPDATE `USING` policy checks the existing row, which already has the correct ownership column value. |
| 17 | + |
| 18 | +### Per-PK savepoint isolation |
| 19 | + |
| 20 | +Each primary key's flush is wrapped in its own savepoint. When RLS denies a write: |
| 21 | + |
| 22 | +1. The database raises an error inside the savepoint |
| 23 | +2. CloudSync rolls back that savepoint, releasing all resources acquired during the failed statement |
| 24 | +3. Processing continues with the next primary key |
| 25 | + |
| 26 | +This means a single payload can contain a mix of allowed and denied rows — allowed rows commit normally, denied rows are silently skipped. The caller receives the total number of column changes processed (including denied ones) rather than an error. |
| 27 | + |
| 28 | +## Quick Setup |
| 29 | + |
| 30 | +Given a table with an ownership column (`user_id`): |
| 31 | + |
| 32 | +```sql |
| 33 | +CREATE TABLE documents ( |
| 34 | + id TEXT PRIMARY KEY NOT NULL, |
| 35 | + user_id UUID, |
| 36 | + title TEXT, |
| 37 | + content TEXT |
| 38 | +); |
| 39 | + |
| 40 | +SELECT cloudsync_init('documents'); |
| 41 | +``` |
| 42 | + |
| 43 | +Enable RLS and create standard policies: |
| 44 | + |
| 45 | +```sql |
| 46 | +ALTER TABLE documents ENABLE ROW LEVEL SECURITY; |
| 47 | + |
| 48 | +CREATE POLICY "select_own" ON documents FOR SELECT |
| 49 | + USING (auth.uid() = user_id); |
| 50 | + |
| 51 | +CREATE POLICY "insert_own" ON documents FOR INSERT |
| 52 | + WITH CHECK (auth.uid() = user_id); |
| 53 | + |
| 54 | +CREATE POLICY "update_own" ON documents FOR UPDATE |
| 55 | + USING (auth.uid() = user_id) |
| 56 | + WITH CHECK (auth.uid() = user_id); |
| 57 | + |
| 58 | +CREATE POLICY "delete_own" ON documents FOR DELETE |
| 59 | + USING (auth.uid() = user_id); |
| 60 | +``` |
| 61 | + |
| 62 | +## Example: Two-User Sync with RLS |
| 63 | + |
| 64 | +This example shows the complete flow of syncing data between two databases where the target enforces RLS. |
| 65 | + |
| 66 | +### Setup |
| 67 | + |
| 68 | +```sql |
| 69 | +-- Source database (DB A) — no RLS, represents the sync server |
| 70 | +CREATE TABLE documents ( |
| 71 | + id TEXT PRIMARY KEY NOT NULL, user_id UUID, title TEXT, content TEXT |
| 72 | +); |
| 73 | +SELECT cloudsync_init('documents'); |
| 74 | + |
| 75 | +-- Target database (DB B) — RLS enforced |
| 76 | +CREATE TABLE documents ( |
| 77 | + id TEXT PRIMARY KEY NOT NULL, user_id UUID, title TEXT, content TEXT |
| 78 | +); |
| 79 | +SELECT cloudsync_init('documents'); |
| 80 | +ALTER TABLE documents ENABLE ROW LEVEL SECURITY; |
| 81 | +-- (policies as above) |
| 82 | +``` |
| 83 | + |
| 84 | +### Insert sync |
| 85 | + |
| 86 | +User 1 creates a document on DB A: |
| 87 | + |
| 88 | +```sql |
| 89 | +-- On DB A |
| 90 | +INSERT INTO documents VALUES ('doc1', 'user1-uuid', 'Hello', 'World'); |
| 91 | +``` |
| 92 | + |
| 93 | +Apply the payload on DB B as the authenticated user: |
| 94 | + |
| 95 | +```sql |
| 96 | +-- On DB B (running as user1) |
| 97 | +SET app.current_user_id = 'user1-uuid'; |
| 98 | +SET ROLE authenticated; |
| 99 | +SELECT cloudsync_payload_apply(decode(:payload_hex, 'hex')); |
| 100 | +``` |
| 101 | + |
| 102 | +The insert succeeds because `user_id` matches `auth.uid()`. |
| 103 | + |
| 104 | +### Insert denial |
| 105 | + |
| 106 | +User 1 tries to sync a document owned by user 2: |
| 107 | + |
| 108 | +```sql |
| 109 | +-- On DB A |
| 110 | +INSERT INTO documents VALUES ('doc2', 'user2-uuid', 'Secret', 'Data'); |
| 111 | +``` |
| 112 | + |
| 113 | +```sql |
| 114 | +-- On DB B (running as user1) |
| 115 | +SET app.current_user_id = 'user1-uuid'; |
| 116 | +SET ROLE authenticated; |
| 117 | +SELECT cloudsync_payload_apply(decode(:payload_hex, 'hex')); |
| 118 | +``` |
| 119 | + |
| 120 | +The insert is denied by RLS. The row does not appear in DB B. No error is raised to the caller — CloudSync isolates the failure via a per-PK savepoint and continues processing the remaining payload. |
| 121 | + |
| 122 | +### Partial update sync |
| 123 | + |
| 124 | +User 1 updates only the title of their own document: |
| 125 | + |
| 126 | +```sql |
| 127 | +-- On DB A |
| 128 | +UPDATE documents SET title = 'Hello Updated' WHERE id = 'doc1'; |
| 129 | +``` |
| 130 | + |
| 131 | +The sync payload contains only the changed column (`title`). CloudSync detects that the row already exists on DB B and uses a plain `UPDATE` statement: |
| 132 | + |
| 133 | +```sql |
| 134 | +UPDATE documents SET title = $2 WHERE id = $1; |
| 135 | +``` |
| 136 | + |
| 137 | +The UPDATE policy checks the existing row (which has the correct `user_id`), so it succeeds. |
| 138 | + |
| 139 | +### Mixed payload |
| 140 | + |
| 141 | +When a single payload contains rows for multiple users, CloudSync handles each primary key independently: |
| 142 | + |
| 143 | +```sql |
| 144 | +-- On DB A |
| 145 | +INSERT INTO documents VALUES ('doc3', 'user1-uuid', 'Mine', '...'); |
| 146 | +INSERT INTO documents VALUES ('doc4', 'user2-uuid', 'Theirs', '...'); |
| 147 | +``` |
| 148 | + |
| 149 | +```sql |
| 150 | +-- On DB B (running as user1) |
| 151 | +SELECT cloudsync_payload_apply(decode(:payload_hex, 'hex')); |
| 152 | +-- doc3 is inserted (allowed), doc4 is silently skipped (denied) |
| 153 | +``` |
| 154 | + |
| 155 | +## Supabase Notes |
| 156 | + |
| 157 | +When using Supabase: |
| 158 | + |
| 159 | +1. **auth.uid()**: Returns the authenticated user's UUID from the JWT claims. |
| 160 | +2. **JWT propagation**: Ensure the JWT token is set before sync operations: |
| 161 | + ```sql |
| 162 | + SELECT set_config('request.jwt.claims', '{"sub": "user-uuid", ...}', true); |
| 163 | + ``` |
| 164 | +3. **Service role bypass**: The Supabase service role bypasses RLS entirely. Use the `authenticated` role for user-context operations where RLS enforcement is desired. |
| 165 | + |
| 166 | +## Troubleshooting |
| 167 | + |
| 168 | +### "new row violates row-level security policy" |
| 169 | + |
| 170 | +**Symptom**: Insert operations fail during sync. |
| 171 | + |
| 172 | +**Cause**: The ownership column value doesn't match the authenticated user. |
| 173 | + |
| 174 | +**Solution**: Verify that: |
| 175 | +- The JWT / session variable is set correctly before calling `cloudsync_payload_apply` |
| 176 | +- The `user_id` column in the synced data matches `auth.uid()` |
| 177 | +- RLS policies reference the correct ownership column |
| 178 | + |
| 179 | +### Debugging |
| 180 | + |
| 181 | +```sql |
| 182 | +-- Check current auth context |
| 183 | +SELECT auth.uid(); |
| 184 | + |
| 185 | +-- Inspect a specific row's ownership |
| 186 | +SELECT id, user_id FROM documents WHERE id = 'problematic-pk'; |
| 187 | + |
| 188 | +-- Temporarily disable RLS to inspect all data |
| 189 | +ALTER TABLE documents DISABLE ROW LEVEL SECURITY; |
| 190 | +-- ... inspect ... |
| 191 | +ALTER TABLE documents ENABLE ROW LEVEL SECURITY; |
| 192 | +``` |
0 commit comments