Postgresql partition and sqlalchemy
Maybe a bit late, but I would like to share what I built upon @moshevi 's and @Seb 's answers:
In my IoT use-case, I required actual sub-partitioning (first level year, second level nodeid). Also I wanted to generalize it slightly.
This is what I came up with:
from sqlalchemy.ext.declarative import DeclarativeMeta
from sqlalchemy.sql.ddl import DDL
from sqlalchemy import event
class PartitionByMeta(DeclarativeMeta):
def __new__(cls, clsname, bases, attrs, *, partition_by, partition_type):
@classmethod
def get_partition_name(cls_, suffix):
return f'{cls_.__tablename__}_{suffix}'
@classmethod
def create_partition(cls_, suffix, partition_stmt, subpartition_by=None, subpartition_type=None):
if suffix not in cls_.partitions:
partition = PartitionByMeta(
f'{clsname}{suffix}',
bases,
{'__tablename__': cls_.get_partition_name(suffix)},
partition_type = subpartition_type,
partition_by=subpartition_by,
)
partition.__table__.add_is_dependent_on(cls_.__table__)
event.listen(
partition.__table__,
'after_create',
DDL(
# For non-year ranges, modify the FROM and TO below
# LIST: IN ('first', 'second');
# RANGE: FROM ('{key}-01-01') TO ('{key+1}-01-01')
f"""
ALTER TABLE {cls_.__tablename__}
ATTACH PARTITION {partition.__tablename__}
{partition_stmt};
"""
)
)
cls_.partitions[suffix] = partition
return cls_.partitions[suffix]
if partition_by is not None:
attrs.update(
{
'__table_args__': attrs.get('__table_args__', ())
+ (dict(postgresql_partition_by=f'{partition_type.upper()}({partition_by})'),),
'partitions': {},
'partitioned_by': partition_by,
'get_partition_name': get_partition_name,
'create_partition': create_partition
}
)
return super().__new__(cls, clsname, bases, attrs)
Which is to be used as follows, assuming the respective VehicleDataMixin class to be created as introduced by @moshevi
class VehicleData(VehicleDataMixin, Project, metaclass=PartitionByMeta, partition_by='timestamp',partition_type='RANGE'):
__tablename__ = 'vehicle_data'
__table_args__ = (
Index('ts_ch_nod_idx', "timestamp", "nodeid", "channelid", postgresql_using='brin'),
UniqueConstraint('timestamp','nodeid','channelid', name='ts_ch_nod_constr')
)
Which can then be subpartitoned iteratively like so (to be adapted)
for y in range(2017, 2021):
# Creating tables for all known nodeids
tbl_vehid_y = VehicleData.create_partition(
f"{y}", partition_stmt=f"""FOR VALUES FROM ('{y}-01-01') TO ('{y+1}-01-01')""",
subpartition_by='nodeid', subpartition_type='LIST'
)
for i in {3, 4, 7, 9}:
# Creating all the years below these nodeids including a default partition
tbl_vehid_y.create_partition(
f"nid{i}", partition_stmt=f"""FOR VALUES IN ('{i}')"""
)
# Defaults (nodeid) per year partition
tbl_vehid_y.create_partition("def", partition_stmt="DEFAULT")
# Default to any other year than anticipated
VehicleData.create_partition("def", partition_stmt="DEFAULT")
partition_by='timestamp' <= This is the column to partition by
partition_type='RANGE' <= This is the (PSQL specific) partition type
partition_stmt=f"""FOR VALUES IN ('{i}')""" <= This is the (PSQL specific) partitioning statement.
I had a similar problem. I found @moshevi's answer quite useful, and ended up generalising it a bit (as I had many tables to partition).
First, create a metaclass such as this:
from sqlalchemy.ext.declarative import DeclarativeMeta
from sqlalchemy.sql.ddl import DDL
from sqlalchemy import event
class PartitionByYearMeta(DeclarativeMeta):
def __new__(cls, clsname, bases, attrs, *, partition_by):
@classmethod
def get_partition_name(cls_, key):
# 'measures' -> 'measures_2020' (customise as needed)
return f'{cls_.__tablename__}_{key}'
@classmethod
def create_partition(cls_, key):
if key not in cls_.partitions:
Partition = type(
f'{clsname}{key}', # Class name, only used internally
bases,
{'__tablename__': cls_.get_partition_name(key)}
)
Partition.__table__.add_is_dependent_on(cls_.__table__)
event.listen(
Partition.__table__,
'after_create',
DDL(
# For non-year ranges, modify the FROM and TO below
f"""
ALTER TABLE {cls_.__tablename__}
ATTACH PARTITION {Partition.__tablename__}
FOR VALUES FROM ('{key}-01-01') TO ('{key+1}-01-01');
"""
)
)
cls_.partitions[key] = Partition
return cls_.partitions[key]
attrs.update(
{
# For non-RANGE partitions, modify the `postgresql_partition_by` key below
'__table_args__': attrs.get('__table_args__', ())
+ (dict(postgresql_partition_by=f'RANGE({partition_by})'),),
'partitions': {},
'partitioned_by': partition_by,
'get_partition_name': get_partition_name,
'create_partition': create_partition
}
)
return super().__new__(cls, clsname, bases, attrs)
Next, for any table in your model that you want to partition:
class MeasureMixin:
# The columns need to be pulled out into this mixin
# Note: any foreign key columns will need to be wrapped like this:
@declared_attr
def city_id(self):
return Column(ForeignKey('cities.id'), not_null=True)
log_date = Column(Date, not_null=True)
peaktemp = Column(Integer)
unitsales = Column(Integer)
class Measure(MeasureMixin, Base, metaclass=PartitionByYearMeta, partition_by='logdate'):
__tablename__ = 'measures'
This makes it easy to add more tables and partition by any number of values.
Creating a new partition on the fly works like this:
# Make sure you commit any session that is currently open, even for select queries:
session.commit()
Partition = Measure.create_partition(2020)
if not engine.dialect.has_table(Partition.__table__.name):
Partition.__table__.create(bind=engine)
Now the partition for key 2020 is created and values for that year can be inserted.
You can use a MeasureMixin that both classes can inherit from.
And then use an event to attach the table partition.
from sqlalchemy import event
class MeasureMixin:
city_id = Column(Integer, not_null=True)
log_date = Column(Date, not_null=True)
peaktemp = Column(Integer)
unitsales = Column(Integer)
class Measure(MeasureMixin, Base):
__tablename__ = 'measures'
__table_args__ = {
postgresql_partition_by: 'RANGE (log_date)'
}
class Measure2020(MeasureMixin, Base):
__tablename__ = 'measures2020'
Measure2020.__table__.add_is_dependent_on(Measure.__table__)
event.listen(
Measure2020.__table__,
"after_create",
DDL("""ALTER TABLE measures ATTACH PARTITION measures2020
VALUES FROM ('2020-01-01') TO ('2021-01-01');""")
)