Efficient Django QuerySet Use

Efﬁcient Django
QuerySet Use
Jayson Falkner, PhD
jayson@counsyl.com
@jaysonfalkner

Overview
• 30 min talk
• How to go from Django QuerySet to SQL
• Straight-forward way to proﬁle
• Fast CRUD (Create, Read, Update, Delete)

Assumptions
• You are using Python + Django
• Avoiding direct SQL use
• Postgres (aka PostgreSQL)
• Performance matters
• Slides and code are on GitHub
github.com/jfalkner/Efﬁcient-Django-QuerySet-Use

A simple data model
class Sample(models.Model):
barcode = models.CharField(max_length=10, unique=True)
production = models.BooleanField()
created = models.DateTimeField()
def status(self):
return self.statuses.all()[0]
class SampleStatus(models.Model):
sample = models.ForeignKey(Sample,
related_name='statuses')
status_code = models.PositiveSmallIntegerField()
RECEIVED = 1; LAB = 2; COMPLETE = 3

Don’t do this
# What samples are in the lab?
samples = Sample.objects.filter(
production=True,
statuses__status_code=SampleStatus.LAB)
# Loop through. It’s business time!
for s in samples:
do_something(s.barcode,
s.status().created)

What takes time?
Perform SQL
Serialize ResultsParse SQL
Runtime (Django +Your Code)
Parse ResultsSerialize SQL
Python
Postgres (RDMS)
Socket

Poor Man’s Django/QuerySet Proﬁling
from datetime import datetime
from django.db import connection as con
from example.models import Sample
start = datetime.now()
query_count = len(con.queries)
production=True,
statuses__status_code =SampleStatus.LAB)
for s in samples:
s.status().created)
print “Time: %s”%(datetime.now()-start)
print “Queries: %s”%(len(con.queries)-query_count)
for query in connection.queries[query_count:]:
print query

from example.models import Sample, SampleStatus
start = datetime.now()
query_count = len(con.queries)
production=True,
statuses__status_code = SampleStatus.LAB)
for s in samples:
s.status().created)
print “Time: %s”%(datetime.now()-start)
print “Queries: %s”%(len(con.queries)-query_count)
for query in connection.queries[query_count:]:
print query

from example.models import Sample, SampleStatus
import django_pg_utils
django_pg_utils.track_sql()
production=True,
statuses__status_code = SampleStatus.LAB)
for s in samples:
s.status().created)
django_pg_utils.print_sql()

# Read all matching samples.
production=True,
# Loop with query for `created` value.
for s in samples:
do_something(s.barcode, s.status().created)
SQL Read in QuerySet (1,000 samples)
Python:1.112s, Postgres: 0.009s, Queries: 1,001
(And one more for each other sample...)
SELECT "id", "sample_id", "status", "created"
FROM "db_samplestatus"
WHERE "sample_id" = 10
ORDER BY "created" DESC LIMIT 1
SELECT "id", "barcode", "production", "created"
FROM "db_sample"
WHERE "db_sample"."production" = TRUE

Why is it slow?
• O(n) queries when it could be O(1)
• One query for all samples
• One query per status of sample
• Serializes/deserializes unused values
• Sample: id, created
• SampleStatus: id, status, sample_id

production=True,
# Cache all lookups once in Python dict.
samples = samples.prefetch_related(‘statuses’)
for s in samples:
do_something(s.barcode, s.status().created)
Python: 0.460s, Postgres: 0.018s, Queries: 2
FROM "db_sample"
INNER JOIN "db_samplestatus" ON
("id" = "db_samplestatus"."sample_id")
WHERE ("production" = true AND
"db_samplestatus"."status_code" = 2 )
SELECT "id", "sample_id", "status_code", "created"
FROM "db_samplestatus"
WHERE "sample_id" IN (1220001, 1220002, 1220003 ...

Improving even more
• O(1) query = Good
• Python dict for lookups is relatively slow
• Let Postgres do the work
• Return only params of interest
• Minimizes serialize/deserialize time
• Smaller memory footprint

Sample.objects
# JOIN and calc max status code from SampleStatus.
.annotate(latest_status_code=Max(
'statuses__status_code'))
.filter(production=True,
latest_status_code__eq=SampleStatus.LAB)
# Return exactly the fields to use.
.values_list('barcode', 'statuses__created')
SELECT "barcode", "created"
FROM "db_sample"
LEFT OUTER JOIN "db_samplestatus" ON
("id" = "db_samplestatus"."sample_id")
WHERE ("production" = TRUE)
GROUP BY "id", "barcode", "production", "created",
"status_code", "db_samplestatus"."status_code"
HAVING MAX("db_samplestatus"."status_code") = 2

latest_status = models.ForeignKey(
'SampleStatus', related_name='+', null=True)
def status(self):
...
Using select_related()
JOIN via denormalized foreign key on Sample

production=True,
samples = samples.select_related('latest_status')
for s in samples:
do_something((s.barcode, s.latest_status.created))
SELECT "id", "barcode", "production", "created",
"latest_status_id",
T3."id", T3."sample_id", T3."status",
T3."status_code", T3."created"
FROM "example_sample"
INNER JOIN "example_samplestatus" ON
("id" = "example_samplestatus"."sample_id")
LEFT OUTER JOIN "example_samplestatus" T3 ON
("latest_status_id" = T3."id")
WHERE ("production" = TRUE
AND "example_samplestatus"."status_code" = 2)

status_code = models.PositiveSmallIntegerField()
status_changed = models.DateTimeField()
def status(self):
...
Denormalize for speed
Change the data model to avoid a JOIN

# Read all at once and return a list of tuples.
Sample.objects
.filter(production=True,
status_code = SampleStatus.LAB)
.values_list('barcode', 'status_changed')
SQL Read in QuerySet (1,000,000 samples)
SELECT "db_sample"."barcode",
"db_sample"."status_code"
FROM "db_sample"
WHERE ("db_sample"."status_code" = 2
AND "db_sample"."production" = TRUE)

Why is it faster?
• JOINs and table scans are relatively slow
• SQL’s EXPLAIN helps show this
postgresql.org/docs/9.1/static/using-explain.html
EXPLAIN SELECT "db_sample"."barcode",
"db_sample"."status_code"
FROM "db_sample"
WHERE ("db_sample"."status_code" = 2
AND "db_sample"."production" = TRUE)
Seq Scan on db_sample (cost=0.00..27203.00 rows=5567 width=8)
Filter: (production AND (status_code = 2))

GroupAggregate (cost=691622.75..787997.34 rows=2965372 width=23)
Filter: (max(db_samplestatus.status_code) = 2)
-> Sort (cost=691622.75..699036.18 rows=2965372 width=23)
Sort Key: id, barcode, production, created, status_code,
-> Hash Right Join (cost=43063.00..190400.49 rows=2965372 width=23)
Hash Cond: (db_samplestatus.sample_id = db_sample.id)
-> Seq Scan on db_samplestatus (cost=0.00..51015.90 rows=2993590 width=6)
-> Hash (cost=24703.00..24703.00 rows=1000000 width=21)
-> Seq Scan on db_sample (cost=0.00..24703.00 rows=1000000 width=21)
Filter: production
SQL Read in QuerySet (1,000,000 samples)
Denormalized = 0.175s
Seq Scan on db_sample (cost=0.00..27203.00 rows=5567 width=8)
Filter: (production AND (status_code = 2))
Sample, SampleStatus JOIN = 11s
Denormalized + Multicolumn INDEX = 0.004s
Index Scan using db_sample_prod_lab on db_sample (cost=0.00..9101.15 rows=5567 width=8)
Index Cond: ((status_code = 2) AND (production = true))
Filter: production

status_code = models.PositiveSmallIntegerField(index=True)
status_changed = models.DateTimeField()
Multi-Column Indexing
from django_pg_utils import pg_multicolumn_index
pg_multicolumn_index(Sample, ['production', 'status_code'])
Django Model ‘index’ won’t help
Postgres-speciﬁc CREATE INDEX helper method

How long does it take?
QuerySet SELECT # Samples Time (Postgres)
Loop 1,000 1.11s (0.01s)
Loop prefetch_selected() 1,000 0.46s (0.02s)
Loop select_related() 10,000 0.71s (0.13s)
values_list() + JOIN 10,000 0.23s (0.19s)
Denorm + INDEX 100,000,000 0.99s (0.38s)
On a MacBook Pro Retina. Postgres 9.1. No conﬁg tweaks.

In context at Counsyl
• How much is 100,000,000 samples?
• query: samples to process this week?
• Consider Counsyl’s main product
• 4,000,000 US pregnancies per year
• Screen 2x people (mother and father)
• That is 12 years of samples!

Oh yeah, other stuff!
• Batching CRUD actions = speed
• Avoid any O(n) loops on model objects
• Bulk CREATE, UPDATE, DELETE exist
• bulk_create(), update(), delete()
• save() won’t work
• Helper code for multi-value UPDATE
docs.djangoproject.com/en/dev/topics/db/queries/

FROM "db_sample"
WHERE "production" = TRUE
SELECT (1) AS "a"
FROM "db_sample"
WHERE "id" = 1 LIMIT 1
samples = Sample.objects.filter(production=True)
# Loop, update, and invoke `save()` for each.
for s in samples:
s.barcode = 'PREFIX'+s.barcode
s.save()
Inefﬁcient UPDATE in QuerySet
Queries: O(n)
UPDATE "db_sample"
SET "barcode" = 'PREFIX0',
"production" = TRUE, "created" = '2013-09-29'
WHERE "db_sample"."id" = 1

# Loop and make samples one at a time.
now = datetime.now()
(Sample.objects.filter(production=True)
.update(created=now))
Efﬁcient UPDATE in QuerySet
UPDATE "db_sample"
SET "created" = '2013-09-29 03:40:18.925695-05:00'
WHERE "db_sample"."production" = TRUE
Queries: O(1)

No multi-value update?
• Django’s update() is limited to one value
• SQL has no such restriction

SET barcode = input.update
FROM
(SELECT unnest(ARRAY[1, 2, 3, 4, 5, 6]),
unnest(ARRAY[6, 6, 6, 6, 6, 6])) AS INPUT (filter,
UPDATE)
WHERE id = INPUT.filter;
values = Sample.objects.values_list('id', 'barcode')
filter_vals = [id for id, _ in values]
update_vals = ['PREFIX'+bar for _, bar in values]
django_pg_utils.pg_bulk_update(Sample, 'id', 'barcode',
filter_vals, update_vals)
Efﬁcient UPDATE in QuerySet
Queries: O(1)

for barcode in barcodes:
Sample.objects.create(barcode=barcode,
production=True,
created=datetime.now())
Inefﬁcient CREATE in QuerySet
Queries: O(n)
INSERT INTO "example_sample" ("barcode",
"production",
"created")
VALUES ('01234', TRUE, '2013-10-04 01:55:15.456163-05:00',
2,
'2013-10-04 01:55:15.456173-05:00')
RETURNING "example_sample"."id"

# Buffer all Sample instances with values in a list.
samples = []
for barcode in barcodes:
samples.append(Sample(barcode=barcode,
production=True,
created=datetime.now())
# Bulk create all samples at once.
Sample.objects.bulk_create(samples)
Efﬁcient CREATE in QuerySet
Queries: O(1)
INSERT INTO "example_sample" ("barcode",
"production",
"created")
VALUES ('100000', TRUE, '2013-10-04 02:00:01.556393-05:00'),
('100001', TRUE, '2013-10-04 02:00:01.556483-05:00'),
('100002', TRUE, '2013-10-04 02:00:01.556526-05:00'),
...

Summary
• Batch/Bulk Everything FTW
• READ: values() and values_list()
•pg_multicolumn_index()
• CREATE: bulk_create()
• UPDATE: update() and pg_bulk_update()
• DELETE: delete()

Efﬁcient Django
QuerySet Use
Jayson Falkner, PhD
jayson@counsyl.com
@jaysonfalkner

Efficient Django QuerySet Use

Recommended

Recommended

More Related Content

Recently uploaded

Recently uploaded (20)

Featured

Featured (20)

Efficient Django QuerySet Use