Tutorial: Backup and Maintenance

This tutorial covers operational calls for durability, backups, restores, soft deletes, compaction, and monitoring.

Flush, checkpoint, and close

collection.flush()
collection.checkpoint()
collection.close()
client.close()

Use commit() after normal write batches:

with collection.insert_session() as session:
    session.add_item([0.1, 0.2, 0.3, 0.4], id=1)

# committed automatically by the session

Use checkpoint() before backups or controlled shutdowns.

Operational meaning:

Call	Purpose
commit()	Commit normal writes.
flush()	Flush pending buffers and bytes without clearing the WAL.
checkpoint()	Force a durable checkpoint and clear committed WAL state.
close()	Flush and close the handle from the API perspective.

For scripts, insert_session() is usually enough. For services and backup jobs, call checkpoint() before taking snapshots or stopping the process.

Collection snapshots

Create a collection snapshot:

db.snapshot_collection("items", "./items.snapshot")

Restore it into a collection:

db.restore_collection(
    "items_restored",
    "./items.snapshot",
    overwrite=True,
)

You can also snapshot from the collection handle:

collection.snapshot_to("./items.snapshot")

Snapshot paths are interpreted where the operation runs:

• local mode: path is on the Python process filesystem;
• remote mode: path is on the server filesystem.

Use snapshots when you want a compact operational backup for the same LynseDB version and environment.

Database snapshots

client.snapshot_database("app", "./app.snapshot")
client.restore_database("app_restored", "./app.snapshot", overwrite=True)

Database-level snapshots are useful for complete environment backups.

Restore into a new name first when testing:

client.restore_database("app_restore_test", "./app.snapshot", overwrite=True)
test_db = client.get_database("app_restore_test")
print(test_db.show_collections())

This validates the backup without replacing production data.

Export and import

Exports are filesystem directories containing JSONL metadata and binary vector payloads.

db.export_collection("items", "./items-export")
db.import_collection("items_copy", "./items-export", overwrite=True)

Use export/import when you want a portable representation instead of a snapshot.

Collection handle shortcuts:

collection.export_to("./items-export")

Export/import is useful for:

• moving data between environments;
• inspecting metadata outside LynseDB;
• long-term portable backups;
• migration scripts.

Snapshot/restore is usually faster for operational backups. Export/import is usually easier to inspect and transform.

Soft delete and restore

Deletes are logical tombstones:

collection.delete_items([10, 11])
print(collection.list_deleted_ids())

collection.restore_items([10])
print(collection.list_deleted_ids())

Soft-deleted IDs are excluded from search and query results.

Deletion lifecycle:

1. delete_items() tombstones rows.
2. Search and query exclude tombstoned IDs.
3. restore_items() can bring tombstoned IDs back.
4. compact() physically removes tombstoned rows.

After compaction, removed rows cannot be restored from the collection itself. Use a snapshot or export if you need recovery after compaction.

Compaction

Compaction physically removes tombstoned vectors and rebuilds storage.

removed = collection.compact()
print(f"removed {removed} rows")

Run compaction during a maintenance window for large collections.

Check before and after:

before = collection.stats()
removed = collection.compact()
after = collection.stats()

print(before)
print(removed)
print(after)

Stats and inspection

print(collection.shape)
print(collection.stats())
print(collection.index_mode)
print(collection.list_fields())
print(collection.list_vector_fields())

Database-level inspection:

print(client.list_databases())
print(db.show_collections())
print(db.show_collections_details())

show_collections_details() returns a pandas DataFrame if pandas is installed; otherwise it returns a list of dictionaries.

Field and vector-field inspection:

print(collection.list_fields())
print(collection.list_vector_fields())
print(collection.max_id)
print(collection.is_id_exists(123))

HTTP operations

For remote deployments:

curl http://127.0.0.1:7637/healthz
curl http://127.0.0.1:7637/readyz
curl http://127.0.0.1:7637/metrics
curl http://127.0.0.1:7637/openapi.json

With authentication:

curl -H "Authorization: Bearer your_key" http://127.0.0.1:7637/list_databases

Important metrics include:

• request counts and request latency;
• WAL bytes;
• data directory bytes;
• vector index bytes;
• process memory;
• index build started, completed, failed, in-progress, and duration values.

Use /openapi.json to inspect the exact server schema for automation or custom clients.

Backup recipes

Local collection backup

collection.checkpoint()
collection.snapshot_to("./backups/items.snapshot")

Local database backup

for name in client.list_databases():
    client.snapshot_database(name, f"./backups/{name}.snapshot")

Remote backup

client = lynse.VectorDBClient("http://127.0.0.1:7637", api_key="your_key")
client.snapshot_database("app", "/backups/app.snapshot")

Remember that remote paths are server paths.

Restore drill

client.restore_database("app_drill", "./backups/app.snapshot", overwrite=True)
drill_db = client.get_database("app_drill")
print(drill_db.show_collections_details())

Run restore drills regularly. A backup is only useful after you have verified that it can be restored.

Migration recipe

Use export/import when you want to transform or move collection data:

db.export_collection("items", "./items-export")
db.import_collection("items_copy", "./items-export", overwrite=True)

After import:

copied = db.get_collection("items_copy")
print(copied.shape)
print(copied.stats())

Rebuild indexes if your migration changes index strategy:

copied.build_index("HNSW-L2")
copied.checkpoint()

Production checklist

• Use HTTP mode for multi-process access.
• Pin the LynseDB package and server image versions.
• Keep /healthz, /readyz, and /metrics monitored.
• Use API keys when the server is reachable outside a trusted network.
• Call checkpoint() before snapshots.
• Test restore procedures before relying on backups.
• Compare ANN recall against a flat baseline before changing index settings.
• Use persistent volumes for Docker and Kubernetes deployments.
• Set request, payload, top-k, batch, and collection-size limits.
• Keep an index rebuild plan for large ingestion or migration jobs.
• Compact tombstoned rows during planned maintenance windows.
• Keep old snapshots until new snapshots have passed a restore drill.