diff --git a/.gitignore b/.gitignore
index d8a1bb0..38411fc 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,2 +1,3 @@
 __pycache__
 .coverage
+.mypy_cache
diff --git a/backup/backup.py b/backup/backup.py
new file mode 100755
index 0000000..2c0bbfa
--- /dev/null
+++ b/backup/backup.py
@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+from pyln.client import Plugin
+from pprint import pprint
+from collections import namedtuple
+from urllib.parse import urlparse
+import struct
+import os
+from typing import Mapping, Type, Optional
+import logging
+import sys
+from binascii import hexlify
+
+
+plugin = Plugin()
+
+root = logging.getLogger()
+root.setLevel(logging.DEBUG)
+
+handler = logging.StreamHandler(sys.stdout)
+handler.setLevel(logging.DEBUG)
+formatter = logging.Formatter('%(message)s')
+handler.setFormatter(formatter)
+root.addHandler(handler)
+
+# A change that was proposed by c-lightning that needs saving to the
+# backup. `version` is the database version before the transaction was
+# applied.
+Change = namedtuple('Change',['version', 'transaction'])
+
+class Backend(object):
+    def __init__(self, destination: str):
+        raise NotImplementedError
+
+    def snapshot(self, filename: str) -> bool:
+        raise NotImplementedError
+
+    def add_change(self, change: Change) -> bool:
+        raise NotImplementedError
+
+    def initialize(self) -> bool:
+        raise NotImplementedError
+
+class FileBackend(Backend):
+    def __init__(self, destination: str):
+        self.version = None
+        self.prev_version = None
+        self.destination = destination
+        self.offsets = [0, 0]
+        self.version_count = 0
+        self.url = urlparse(self.destination)
+
+    def initialize(self) -> bool:
+        if not os.path.exists(self.url.path):
+            return False
+        return self.read_metadata()
+
+    def write_metadata(self):
+        blob = struct.pack("!IIQIQQ", 0x01, self.version, self.offsets[0],
+                           self.prev_version, self.offsets[1],
+                           self.version_count)
+
+        # Pad the header
+        blob += b'\x00' * (512 - len(blob))
+        mode = "rb+" if os.path.exists(self.url.path) else "wb+"
+
+        with open(self.url.path, mode) as f:
+            f.seek(0)
+            f.write(blob)
+            f.flush()
+
+    def read_metadata(self):
+        with open(self.url.path, 'rb') as f:
+            blob = f.read(512)
+            if len(blob) != 512:
+                logging.warn("Corrupt FileBackend header, expected 512 bytes, got {} bytes".format(len(blob)))
+                return False
+
+            file_version, = struct.unpack_from("!I", blob)
+            if file_version != 1:
+                logging.warn("Unknown FileBackend version {}".format(file_version))
+                return False
+
+            self.version, self.offsets[0], self.prev_version, self.offsets[1], self.version_count, = struct.unpack_from("!IQIQQ", blob, offset=4)
+
+        return True
+
+    def add_change(self, entry: Change) -> bool:
+        typ = b'\x01' if entry.snapshot is None else b'\x02'
+        if typ == b'\x01':
+            payload = b'\x00'.join([t.encode('ASCII') for t in entry.transaction])
+        elif typ == b'\x02':
+            payload = entry.snapshot
+
+        length = struct.pack("!I", len(payload))
+        with open(self.url.path, 'ab') as f:
+            f.seek(self.offsets[0])
+            f.write(length)
+            f.write(payload)
+            self.prev_version, self.offsets[1] = self.version, self.offsets[0]
+            self.version = entry.version
+            self.offsets[0] += 4 + len(payload)
+        self.write_metadata()
+
+        return True
+
+    def rewind(self):
+        # After rewinding we set offsets[0] and prev_version to 0 (best effort
+        # result). If either of these are set to 0 we have two consecutive
+        # rewinds which cannot be safely done (we'd be rewinding more than the
+        # one in-flight transaction).
+        if self.offsets[1] == 0 or self.prev_version == 0:
+            logging.warn("Cannot rewind multiple times.")
+            return False
+
+        self.version, self.offsets[0] = self.prev_version, self.offsets[1]
+        self.prev_version, self.offsets[1] = 0, 0
+        return True
+
+backend_map: Mapping[str, Type[Backend]] = {
+    'file': FileBackend,
+}
+
+def abort(reason: str) -> None:
+    plugin.log(reason)
+    plugin.rpc.stop()
+    raise ValueError()
+
+
+def check_first_write(plugin, data_version):
+    """Verify that we are up-to-date and c-lightning didn't forget writes.
+
+    We may be at most 1 write off:
+
+     - c-lightning and backup are at the same version (happy case)
+     - c-lightning is 1 write behind: it must've crashed inbetween calling the
+       hook and committing the DB transaction.
+     - c-lightning is one or more writes ahead: either we lost some writes, or
+       c-lightning was running without the plugin at some point -> crash!
+     - c-lighning is more than 1 write behind: c-lightning had a lobotomy, or
+       was restored from an old backup -> crash!
+    """
+    backend = plugin.backend
+
+    logging.info("Comparing backup version {} versus first write version {}".format(
+        backend.version, data_version
+    ))
+
+    if backend.version  == data_version - 1:
+        logging.info("Versions match up")
+        return True
+
+    elif backend.prev_version == data_version - 1 and plugin.backend.rewind():
+        logging.info("Last changes not applied, rewinding non-committed transaction")
+        return True
+
+    elif backend.prev_version > data_version - 1:
+        abort("c-lightning seems to have lost some state (failed restore?). Emergency shutdown.")
+
+    else:
+        abort("Backup is out of date, we cannot continue safely. Emergency shutdown.")
+
+
+@plugin.hook('db_write')
+def on_db_write(writes, data_version, plugin, **kwargs):
+    change = Change(data_version, writes)
+    if not hasattr(plugin, 'backend'):
+        plugin.early_writes.append(change)
+        return True
+    else:
+        return apply_write(plugin, change)
+
+
+def apply_write(plugin, change):
+    if not plugin.initialized:
+        assert(check_first_write(plugin, change.version))
+        plugin.initialized = True
+
+    return plugin.backend.add_entry(change)
+
+
+@plugin.init()
+def on_init(options: Mapping[str, str], plugin: Plugin, **kwargs):
+    # Reach into the DB and
+    configs = plugin.rpc.listconfigs()
+    plugin.db_path = configs['wallet']
+    destination = options['backup-destination']
+
+    if not plugin.db_path.startswith('sqlite3'):
+        abort("The backup plugin only works with the sqlite3 database.")
+
+    if destination == 'null':
+        abort("You must specify a backup destination, possibly on a secondary disk.")
+
+    # Let's initialize the backed. First we need to figure out which backend to use.
+    p = urlparse(destination)
+    backend_cl = backend_map.get(p.scheme, None)
+    if backend_cl is None:
+        abort("Could not find a backend for scheme {p.scheme}".format(p=p))
+
+    plugin.backend = backend_cl(destination)
+    if not plugin.backend.initialize():
+        abort("Could not initialize the backup {}, please use 'backup-cli' to initialize the backup first.".format(destination))
+
+    for c in plugin.early_writes:
+        apply_write(plugin, c)
+
+
+plugin.add_option(
+    'backup-destination', None,
+    'Destination of the database backups (file:///filename/on/another/disk/).'
+)
+
+
+if __name__ == "__main__":
+    # Did we perform the version check of backend versus the first write?
+    plugin.initialized = False
+    plugin.early_writes = []
+    plugin.run()
diff --git a/backup/fixtures/backup.dbak b/backup/fixtures/backup.dbak
new file mode 100644
index 0000000..06b3e4e
Binary files /dev/null and b/backup/fixtures/backup.dbak differ
diff --git a/backup/test_backup.py b/backup/test_backup.py
new file mode 100644
index 0000000..db2bf19
--- /dev/null
+++ b/backup/test_backup.py
@@ -0,0 +1,135 @@
+from flaky import flaky
+from pyln.client import RpcError
+from pyln.testing.fixtures import *
+import os
+import time
+
+plugin_dir = os.path.dirname(__file__)
+plugin_path =  os.path.join(plugin_dir, "backup.py")
+
+
+def test_start(node_factory, directory):
+    opts = {
+        'plugin': plugin_path,
+        'backup-destination': 'file://' + os.path.join(directory, 'backup.dbak')
+        }
+    shutil.copyfile(
+        os.path.join(plugin_dir, 'fixtures', "backup.dbak"),
+        os.path.join(directory, "backup.dbak")
+    )
+    l1 = node_factory.get_node(options=opts)
+
+    l1.daemon.wait_for_log(r'backup.py')
+
+    # Restart the node a couple of times, to check that we can resume normally
+    for i in range(5):
+        l1.restart()
+        l1.daemon.wait_for_log(r'Versions match up')
+
+
+def test_tx_abort(node_factory, directory):
+    """Simulate a crash between hook call and DB commit.
+
+    We simulate this by updating the data_version var in the database before
+    restarting the node. This desyncs the node from the backup, and restoring
+    may not work (depending on which transaction was pretend-rolled-back), but
+    continuing should work fine, since it can happen that we crash just
+    inbetween the hook call and the DB transaction.
+
+    """
+    opts = {
+        'plugin': plugin_path,
+        'backup-destination': 'file://' + os.path.join(directory, 'backup.dbak')
+        }
+    shutil.copyfile(
+        os.path.join(plugin_dir, 'fixtures', "backup.dbak"),
+        os.path.join(directory, "backup.dbak")
+    )
+    l1 = node_factory.get_node(options=opts)
+    l1.stop()
+
+    print(l1.db.query("SELECT * FROM vars;"))
+
+    # Now fudge the data_version:
+    l1.db.execute("UPDATE vars SET intval = intval - 1 WHERE name = 'data_version'")
+
+    print(l1.db.query("SELECT * FROM vars;"))
+
+    l1.restart()
+    l1.daemon.wait_for_log(r'Last changes not applied')
+
+
+@flaky
+def test_failing_restore(nf, directory):
+    """The node database is having memory loss, make sure we abort.
+
+    We simulate a loss of transactions by manually resetting the data_version
+    in the database back to n-2, which is non-recoverable.
+
+    """
+    opts = {
+        'plugin': plugin_path,
+        'backup-destination': 'file://' + os.path.join(directory, 'backup.dbak')
+        }
+    shutil.copyfile(
+        os.path.join(plugin_dir, 'fixtures', "backup.dbak"),
+        os.path.join(directory, "backup.dbak")
+    )
+    l1 = node_factory.get_node(options=opts)
+    l1.stop()
+
+    # Now fudge the data_version:
+    l1.db.execute("UPDATE vars SET intval = intval - 2 WHERE name = 'data_version'")
+
+    with pytest.raises(Exception):
+        l1.start()
+
+    l1.daemon.proc.wait()
+    assert(l1.daemon.is_in_log(r'lost some state') is not None)
+
+
+def test_intermittent_backup(node_factory, directory):
+    """Simulate intermittent use of the backup, or an old file backup.
+
+    """
+
+    opts = {
+        'plugin': plugin_path,
+        'backup-destination': 'file://' + os.path.join(directory, 'backup.dbak')
+        }
+    shutil.copyfile(
+        os.path.join(plugin_dir, 'fixtures', "backup.dbak"),
+        os.path.join(directory, "backup.dbak")
+    )
+    l1 = node_factory.get_node(options=opts)
+
+    # Now start without the plugin. This should work fine.
+    del l1.daemon.opts['plugin']
+    del l1.daemon.opts['backup-destination']
+    l1.restart()
+
+    # Now restart adding the plugin again, and it should fail due to gaps in
+    # the backup.
+    l1.stop()
+    with pytest.raises(Exception):
+        l1.daemon.opts.update(opts)
+        l1.start()
+
+    l1.daemon.proc.wait()
+    assert(l1.daemon.is_in_log(r'Backup is out of date') is not None)
+
+
+def test_restore(nf, directory):
+    bpath = os.path.join(directory, 'lightning-1', 'regtest')
+    bdest = 'file://' + os.path.join(bpath, 'backup.dbak')
+    os.makedirs(bpath)
+    subprocess.check_call([cli_path, "init", bpath, bdest])
+    opts = {
+        'plugin': plugin_path,
+        'backup-destination': bdest,
+        }
+    l1 = nf.get_node(options=opts, cleandir=False)
+    l1.stop()
+
+    rdest = os.path.join(bpath, 'lightningd.sqlite.restore')
+    subprocess.check_call([cli_path, "restore", bdest, rdest])