hub/scribe/db/interface.py
Jack Robison dfef80d9c2
initial
2022-03-08 12:41:51 -05:00

273 lines
12 KiB
Python

import struct
import typing
import rocksdb
from typing import Optional
from scribe.db.common import DB_PREFIXES, COLUMN_SETTINGS
from scribe.db.revertable import RevertableOpStack, RevertablePut, RevertableDelete
ROW_TYPES = {}
class PrefixRowType(type):
def __new__(cls, name, bases, kwargs):
klass = super().__new__(cls, name, bases, kwargs)
if name != "PrefixRow":
ROW_TYPES[klass.prefix] = klass
cache_size = klass.cache_size
COLUMN_SETTINGS[klass.prefix] = {
'cache_size': cache_size,
}
return klass
class PrefixRow(metaclass=PrefixRowType):
prefix: bytes
key_struct: struct.Struct
value_struct: struct.Struct
key_part_lambdas = []
cache_size: int = 1024 * 1024 * 64
def __init__(self, db: 'rocksdb.DB', op_stack: RevertableOpStack):
self._db = db
self._op_stack = op_stack
self._column_family = self._db.get_column_family(self.prefix)
if not self._column_family.is_valid:
raise RuntimeError('column family is not valid')
def iterate(self, prefix=None, start=None, stop=None, reverse: bool = False, include_key: bool = True,
include_value: bool = True, fill_cache: bool = True, deserialize_key: bool = True,
deserialize_value: bool = True):
if not prefix and not start and not stop:
prefix = ()
if prefix is not None:
prefix = self.pack_partial_key(*prefix)
if stop is None:
try:
stop = (int.from_bytes(prefix, byteorder='big') + 1).to_bytes(len(prefix), byteorder='big')
except OverflowError:
stop = (int.from_bytes(prefix, byteorder='big') + 1).to_bytes(len(prefix) + 1, byteorder='big')
else:
stop = self.pack_partial_key(*stop)
else:
if start is not None:
start = self.pack_partial_key(*start)
if stop is not None:
stop = self.pack_partial_key(*stop)
if deserialize_key:
key_getter = lambda k: self.unpack_key(k)
else:
key_getter = lambda k: k
if deserialize_value:
value_getter = lambda v: self.unpack_value(v)
else:
value_getter = lambda v: v
it = self._db.iterator(
start or prefix, self._column_family, iterate_lower_bound=(start or prefix),
iterate_upper_bound=stop, reverse=reverse, include_key=include_key,
include_value=include_value, fill_cache=fill_cache, prefix_same_as_start=False
)
if include_key and include_value:
for k, v in it:
yield key_getter(k[1]), value_getter(v)
elif include_key:
for k in it:
yield key_getter(k[1])
elif include_value:
for v in it:
yield value_getter(v)
else:
for _ in it:
yield None
def get(self, *key_args, fill_cache=True, deserialize_value=True):
v = self._db.get((self._column_family, self.pack_key(*key_args)), fill_cache=fill_cache)
if v:
return v if not deserialize_value else self.unpack_value(v)
def get_pending(self, *key_args, fill_cache=True, deserialize_value=True):
packed_key = self.pack_key(*key_args)
last_op = self._op_stack.get_last_op_for_key(packed_key)
if last_op:
if last_op.is_put:
return last_op.value if not deserialize_value else self.unpack_value(last_op.value)
else: # it's a delete
return
v = self._db.get((self._column_family, packed_key), fill_cache=fill_cache)
if v:
return v if not deserialize_value else self.unpack_value(v)
def stage_put(self, key_args=(), value_args=()):
self._op_stack.append_op(RevertablePut(self.pack_key(*key_args), self.pack_value(*value_args)))
def stage_delete(self, key_args=(), value_args=()):
self._op_stack.append_op(RevertableDelete(self.pack_key(*key_args), self.pack_value(*value_args)))
@classmethod
def pack_partial_key(cls, *args) -> bytes:
return cls.prefix + cls.key_part_lambdas[len(args)](*args)
@classmethod
def pack_key(cls, *args) -> bytes:
return cls.prefix + cls.key_struct.pack(*args)
@classmethod
def pack_value(cls, *args) -> bytes:
return cls.value_struct.pack(*args)
@classmethod
def unpack_key(cls, key: bytes):
assert key[:1] == cls.prefix, f"prefix should be {cls.prefix}, got {key[:1]}"
return cls.key_struct.unpack(key[1:])
@classmethod
def unpack_value(cls, data: bytes):
return cls.value_struct.unpack(data)
@classmethod
def unpack_item(cls, key: bytes, value: bytes):
return cls.unpack_key(key), cls.unpack_value(value)
def estimate_num_keys(self) -> int:
return int(self._db.get_property(b'rocksdb.estimate-num-keys', self._column_family).decode())
class BasePrefixDB:
"""
Base class for a revertable rocksdb database (a rocksdb db where each set of applied changes can be undone)
"""
UNDO_KEY_STRUCT = struct.Struct(b'>Q32s')
PARTIAL_UNDO_KEY_STRUCT = struct.Struct(b'>Q')
def __init__(self, path, max_open_files=64, secondary_path='', max_undo_depth: int = 200, unsafe_prefixes=None):
column_family_options = {}
for prefix in DB_PREFIXES:
settings = COLUMN_SETTINGS[prefix.value]
column_family_options[prefix.value] = rocksdb.ColumnFamilyOptions()
column_family_options[prefix.value].table_factory = rocksdb.BlockBasedTableFactory(
block_cache=rocksdb.LRUCache(settings['cache_size']),
)
self.column_families: typing.Dict[bytes, 'rocksdb.ColumnFamilyHandle'] = {}
options = rocksdb.Options(
create_if_missing=True, use_fsync=False, target_file_size_base=33554432,
max_open_files=max_open_files if not secondary_path else -1, create_missing_column_families=True
)
self._db = rocksdb.DB(
path, options, secondary_name=secondary_path, column_families=column_family_options
)
for prefix in DB_PREFIXES:
cf = self._db.get_column_family(prefix.value)
if cf is None and not secondary_path:
self._db.create_column_family(prefix.value, column_family_options[prefix.value])
cf = self._db.get_column_family(prefix.value)
self.column_families[prefix.value] = cf
self._op_stack = RevertableOpStack(self.get, unsafe_prefixes=unsafe_prefixes)
self._max_undo_depth = max_undo_depth
def unsafe_commit(self):
"""
Write staged changes to the database without keeping undo information
Changes written cannot be undone
"""
try:
if not len(self._op_stack):
return
with self._db.write_batch(sync=True) as batch:
batch_put = batch.put
batch_delete = batch.delete
get_column_family = self.column_families.__getitem__
for staged_change in self._op_stack:
column_family = get_column_family(DB_PREFIXES(staged_change.key[:1]).value)
if staged_change.is_put:
batch_put((column_family, staged_change.key), staged_change.value)
else:
batch_delete((column_family, staged_change.key))
finally:
self._op_stack.clear()
def commit(self, height: int, block_hash: bytes):
"""
Write changes for a block height to the database and keep undo information so that the changes can be reverted
"""
undo_ops = self._op_stack.get_undo_ops()
delete_undos = []
if height > self._max_undo_depth:
delete_undos.extend(self._db.iterator(
start=DB_PREFIXES.undo.value + self.PARTIAL_UNDO_KEY_STRUCT.pack(0),
iterate_upper_bound=DB_PREFIXES.undo.value + self.PARTIAL_UNDO_KEY_STRUCT.pack(height - self._max_undo_depth),
include_value=False
))
try:
undo_c_f = self.column_families[DB_PREFIXES.undo.value]
with self._db.write_batch(sync=True) as batch:
batch_put = batch.put
batch_delete = batch.delete
get_column_family = self.column_families.__getitem__
for staged_change in self._op_stack:
column_family = get_column_family(DB_PREFIXES(staged_change.key[:1]).value)
if staged_change.is_put:
batch_put((column_family, staged_change.key), staged_change.value)
else:
batch_delete((column_family, staged_change.key))
for undo_to_delete in delete_undos:
batch_delete((undo_c_f, undo_to_delete))
batch_put((undo_c_f, DB_PREFIXES.undo.value + self.UNDO_KEY_STRUCT.pack(height, block_hash)), undo_ops)
finally:
self._op_stack.clear()
def rollback(self, height: int, block_hash: bytes):
"""
Revert changes for a block height
"""
undo_key = DB_PREFIXES.undo.value + self.UNDO_KEY_STRUCT.pack(height, block_hash)
undo_c_f = self.column_families[DB_PREFIXES.undo.value]
undo_info = self._db.get((undo_c_f, undo_key))
self._op_stack.apply_packed_undo_ops(undo_info)
try:
with self._db.write_batch(sync=True) as batch:
batch_put = batch.put
batch_delete = batch.delete
get_column_family = self.column_families.__getitem__
for staged_change in self._op_stack:
column_family = get_column_family(DB_PREFIXES(staged_change.key[:1]).value)
if staged_change.is_put:
batch_put((column_family, staged_change.key), staged_change.value)
else:
batch_delete((column_family, staged_change.key))
# batch_delete(undo_key)
finally:
self._op_stack.clear()
def get(self, key: bytes, fill_cache: bool = True) -> Optional[bytes]:
cf = self.column_families[key[:1]]
return self._db.get((cf, key), fill_cache=fill_cache)
def iterator(self, start: bytes, column_family: 'rocksdb.ColumnFamilyHandle' = None,
iterate_lower_bound: bytes = None, iterate_upper_bound: bytes = None,
reverse: bool = False, include_key: bool = True, include_value: bool = True,
fill_cache: bool = True, prefix_same_as_start: bool = False, auto_prefix_mode: bool = True):
return self._db.iterator(
start=start, column_family=column_family, iterate_lower_bound=iterate_lower_bound,
iterate_upper_bound=iterate_upper_bound, reverse=reverse, include_key=include_key,
include_value=include_value, fill_cache=fill_cache, prefix_same_as_start=prefix_same_as_start,
auto_prefix_mode=auto_prefix_mode
)
def close(self):
self._db.close()
def try_catch_up_with_primary(self):
self._db.try_catch_up_with_primary()
def stage_raw_put(self, key: bytes, value: bytes):
self._op_stack.append_op(RevertablePut(key, value))
def stage_raw_delete(self, key: bytes, value: bytes):
self._op_stack.append_op(RevertableDelete(key, value))
def estimate_num_keys(self, column_family: 'rocksdb.ColumnFamilyHandle' = None):
return int(self._db.get_property(b'rocksdb.estimate-num-keys', column_family).decode())