PostgreSQL Proficiency
for Python People
Christophe Pettus
PostgreSQL Experts, Inc.
thebuild.com
pgexperts.com
Welcome!
Christophe Pettus
Consultant with PostgreSQL Experts, Inc.
Based in sunny San Francisco, California.
Technical blog: thebuild.com
Twitter: @xof
[email protected]
My background.
PostgreSQL person since 1998.
Came to databases as an application
developer and architect.
Python/Django person since 2008.
What is this?
Just enough PostgreSQL for a Python
developer.
PostgreSQL is a rich environment.
Far too much to learn in a single tutorial.
But enough to be dangerous!
The DevOps World
Integration between development and
operations.
Cross-functional skill sharing.
Maximum automation of development and
deployment processes.
Were way too cheap to hire real
operations staff. Anyway: Cloud!
This means
No experienced DBA on staff.
Have you seen how much those people
cost, anyway?
Development staff pressed into duty as
database administrators.
But its OK its PostgreSQL!
Everyone Loves PostgreSQL!
Fully ACID-compliant relational database
management system.
Richest set of features of any modern
production RDMS.
Relentless focus on quality, security, and
spec compliance.
Capable of very high performance.
PostgreSQL Can Do It.
Tens of thousands of transactions per
second.
Enormous databases (into the petabyte
range).
Support by all Python ORMs and web
frameworks.
Cross-Platform.
Operates natively on all modern operating
systems.
Plus Windows.
Scales from development laptops to huge
enterprise clusters.
Installation
If you have packages
use them!
Provides platform-specific scripting, etc.
RedHat-flavor and Debian-flavor have their
own repositories.
Other OSes have a variety of packaging
systems.
Or you can build from source.
Works on any platform.
Maximum control.
Requires development tools.
Does not come with platform-specific
utility scripts (/etc/init.d, etc.).
Other OSes.
Windows: One-click installer available.
OS X: One-click installer, MacPorts, Fink
and Postgres.app from Heroku.
For other OSes, check postgresql.org.
Creating a database cluster.
A single PostgreSQL server can manage
multiple databases.
The whole group on a single server is
called a cluster.
This is very confusing, yes. Well use the
term server here.
initdb
The command to create a new database is
called initdb.
It creates the files that will hold the
database.
It doesnt automatically start the server.
Many packaging systems automatically
create and start the server for you.
Note on Debian
Debian-style packaging has a sophisticated
cluster management system.
Use it! It will make your life much easier.
pg_createcluster instead of initdb
Just Do This.
Always create databases as UTF-8.
Once created, cannot be changed.
Converting from SQL ASCII to a real
encoding is a total nightmare.
Use your favorite locale, but not C locale.
UTF-8 and system locale are the defaults.
pg_ctl
Built-in command to start and stop
PostgreSQL.
Frequently called by init.d, upstart or other
scripts.
-m fast is the way to go.
Use the package-provided scripts if they
exist; the do the right thing.
psql
Command-line interface to PostgreSQL.
Run queries, examine the schema, look at
PostgreSQLs various views.
Get friendly with it! Its very useful for
doing quick checks.
PostgreSQL directories
All of the data lives under a top-level
directory.
Lets call it $PGDATA.
Find it on your system, and do a ls.
The data lives in base.
The transaction logs live in pg_xlog.
Configuration files.
On most installations, the configuration
files live in $PGDATA.
On Debian-derived systems, they live in
/etc/postgresql/9.3/main/...
Find them.You should see:
postgresql.conf
pg_hba.conf
Configuration
Configuration files.
Only two really matter:
postgresql.conf most server settings.
pg_hba.conf who gets to log in to
what databases?
postgresql.conf
Holds all of the configuration parameters
for the server.
Find it and open it up on your system.
Were All Going To Die.
It Can Be Like This.
Important parameters.
Logging.
Memory.
Checkpoints.
Planner.
Youre done.
No, really, youre done!
Logging.
Be generous with logging; its very lowimpact on the system.
Its your best source of information for
finding performance problems.
Where to log?
syslog If you have a syslog infrastructure
you like already.
standard format to files If you are using
tools that need standard format.
Otherwise, CSV format to files.
What to log?
log_destination = 'csvlog'
log_directory = 'pg_log'
logging_collector = on
log_filename = 'postgres-%Y-%m-%d_%H%M%S'
log_rotation_age = 1d
log_rotation_size = 1GB
log_min_duration_statement = 250ms
log_checkpoints = on
log_connections = on
log_disconnections = on
log_lock_waits = on
log_temp_files = 0
Memory configuration
shared_buffers
work_mem
maintenance_work_mem
shared_buffers
Below 2GB (?), set to 20% of total system
memory.
Below 32GB, set to 25% of total system
memory.
Above 32GB (lucky you!), set to 8GB.
Done.
work_mem
Start low: 32-64MB.
Look for temporary file lines in logs.
Set to 2-3x the largest temp file you see.
Can cause a huge speed-up if set properly!
But be careful: It can use that amount of
memory per planner node.
maintenance_work_mem
10% of system memory, up to1GB.
Maybe even higher if you are having
VACUUM problems.
(Well talk about VACUUM later.)
effective_cache_size
Set to the amount of file system cache
available.
If you dont know, set it to 50% of total
system memory.
And youre done with memory settings.
Checkpoints.
A complete flush of dirty buffers to disk.
Potentially a lot of I/O.
Done when the first of two thresholds are
hit:
A particular number of WAL segments
have been written.
A timeout occurs.
Checkpoint settings.
wal_buffers = 16MB
checkpoint_completion_target = 0.9
checkpoint_timeout = 10m-30m # Depends on restart time
checkpoint_segments = 32 # To start.
Checkpoint settings, 2.
Look for checkpoint entries in the logs.
Happening more often than
checkpoint_timeout?
Adjust checkpoint_segments so that
checkpoints happen due to timeouts
rather filling segments.
And youre done with checkpoint settings.
Checkpoint settings notes.
The WAL can take up to 3 x 16MB x
checkpoint_segments on disk.
Restarting PostgreSQL from a crash can
take up to checkpoint_timeout (but usually
less).
Planner settings.
effective_io_concurrency Set to the
number of I/O channels; otherwise, ignore
it.
random_page_cost 3.0 for a typical
RAID10 array, 2.0 for a SAN, 1.1 for
Amazon EBS.
And youre done with planner settings.
Do not touch.
fsync = on
Never change this.
synchronous_commit = on
Change this, but only if you understand
the data loss potential.
Changing settings.
Most settings just require a server reload
to take effect.
Some require a full server restart (such as
shared_buffers).
Many can be set on a per-session basis!
pg_hba.conf
Users and roles.
A role is a database object that can own
other objects (tables, etc.), and that has
privileges (able to write to a table).
A user is just a role that can log into the
system; otherwise, theyre synonyms.
PostgreSQLs security system is based
around users.
Basic user management.
Dont use the postgres superuser for
anything application-related.
Sadly, you probably will have to grant
schema-modifications privileges to your
application user, if you use migrations.
If you dont have to, dont.
User security.
By default, database traffic is not encrypted.
Turn on ssl if you are running in a cloud
provider.
Have you responded to the Heartbleed
OpenSSL bug, btw?
The WAL.
Why are we talking about this now?
The Write-Ahead Log is key to many
PostgreSQL operations.
Replication, crash recovery, etc., etc.
Dont worry (too much!) about the
internals.
The Basics.
When each transaction is committed, it is
logged to the write-ahead log.
The changes in that transaction are flushed
to disk.
If the system crashes, the WAL is replayed
to bring the database to a consistent state.
A continuous record of changes.
The WAL is a continuous record of changes
since the last checkpoint.
Thus, if you have the disk image of the
database, and every WAL record since that
was created
you can recreate the database to the
end of the WAL.
pg_xlog
The WAL is stored in 16MB segments in
the pg_xlog directory.
Dont mess with it! Never delete anything
out of it!
Records are automatically recycled when
they are no longer required.
WAL archiving.
archive_command
Runs a command each time a WAL
segment is complete.
This command can do whatever you want.
What you want is to move the WAL
segment to someplace safe
on a different system.
On a crash
When PostgreSQL restarts, it replays the
WAL log to bring itself back to a consistent
state.
The WAL segments are essential to proper
crash recovery.
The longer since the last checkpoint, the
more WAL it has to process.
sychronous_commit
When on, COMMIT does not return until
the WAL flush is done.
When off, COMMIT returns when the
WAL flush is queued.
Thus, you might lose transactions on a
crash.
No danger of database corruption.
Backup and
Recovery
pg_dump
Built-in dump/restore tool.
Takes a logical snapshot of the database.
Does not lock the database or prevent
writes to disk.
Low (but not zero) load on the database.
pg_restore
Restores database from a pg_dump.
Is not a fast operation.
Great for simple backups, not suitable for
fast recovery from major failures.
pg_dump / pg_restore advice
Back up globals with pg_dumpall --globalsonly.
Back up each database with pg_dump using
--format=custom.
This allows for a parallel restore using
pg_restore.
pg_restore
Restore using --jobs=<# of cores + 1>.
Most of the time in a restore is spent
rebuilding indexes; this will parallelize that
operation.
Restores are not fast.
PITR backup / recovery
Remember the WAL?
If you take a snapshot of the data
directory
it wont be consistent, but if we add the
WAL records
we can bring it back to consistency.
Getting started with PITR.
Decide where the WAL segments and the
backups will live.
Configure archive_command properly to
do the copying.
Creating a PITR backup.
SELECT pg_start_backup(...);
Copy the disk image and any WAL files that
are created.
SELECT pg_stop_backup();
Make sure you have all the WAL segments.
The disk image + WAL segments are your
backup.
WAL-E
http://github.com/wal-e/wal-e
Provides a full set of appropriate scripting.
Automates create PITR backups into AWS
S3.
Highly recommended!
PITR Restore
Copy the disk image back to where you
need it.
Set up recovery.conf to point to where the
WAL files are.
Start up PostgreSQL, and let it recover.
How long will this take?
The more WAL files, the longer it will take.
Generally takes 10-20% of the time it took
to create the WAL files in the first place.
More frequent snapshots = faster recovery
time.
PITR?
Point-in-time recovery.
You dont have to replay the entire WAL
stream.
It can be stopped at a particular
timestamp, or transaction ID.
Very handy for application-level problems!
Replication.
Hey, what if we sent the WAL directly to
another server?
We could have that server keep up to date
with the primary server!
And thats how PostgreSQL replication
works.
WAL Archiving.
Each 16MB segment is sent to the
secondary when complete.
The secondary reads it, and applies it to its
copy.
Make sure the WAL file copied
automatically.
Use rsync, WAL-E, etc., not scp.
Hmm but what if we
transmitted the WAL changes directly to
the secondary without having to ship the
file?
Great idea!
Such a great idea, PostgreSQL implements
it!
Thats what Streaming Replication is.
Streaming Replication Basics.
The secondary connects via a standard
PostgreSQL connection to the primary.
As changes happen on the primary, they are
sent down to the secondary.
The secondary applies them to its local
copy of the database.
recovery.conf
All replication is orchestrated through the
recovery.conf file.
Always lives in your $PGDATA directory.
Controls how to connect to the primary,
how far to recover (for PITR), etc., etc.
Also used if you are bringing the server up
as a PITR recovery instead of replication.
Disaster recovery.
Always have a disaster recovery strategy.
What if you data center / AWS region goes
down?
Have a plan for recovery from a remote
site.
WAL archiving is a great way to handle this.
pg_basebackup
Utility for doing a snapshot of a running
server.
Easiest way to take a snapshot to start a
new secondary.
Can also be used as an archival backup.
Lets see!
Replication!
Replication, the good.
Easy to set up.
Schema changs are automatically replicated.
Secondary can be used to handle read-only
queries for load balancing.
Very few gotchas; it either works or it
doesnt, and it is vocal about not working.
Replication, the bad.
Entire database or none of it.
No writes of any kind to the secondary.
This includes temporary tables.
Some things arent replicated.
Temporary tables, unlogged tables.
Advice?
Start with WAL-E.
The README tells you everything you
need to know.
Handles a very large number of complex
replication problems easily.
As you scale out of it, youll have the
relevant experience.
Trigger-based replication
Installs triggers on tables on master.
A daemon process picks up the changes
and applies them to the secondaries.
Third-party add-ons to PostgreSQL.
Trigger-based rep: Good.
Highly configurable.
Can push part or all of the tables; dont
have to replicate everything.
Multi-master setups possible (Bucardo).
Trigger-based rep: The bad.
Fiddly and complex to set up.
Schema changes must be pushed out
manually.
Imposes overhead on the master.
Transactions,
MVCC and
VACUUM
Transaction
A unit of which which must be:
Applied atomically to the database.
Invisible to other database clients until it
is committed.
The Classic Example.
BEGIN;
INSERT INTO transactions(account_id, value, offset_id)
VALUES (11, 120.00, 14);
INSERT INTO transactions(account_id, value, offset_id)
VALUES (14, -120.00, 11);
COMMIT;
Transaction Properties.
Once the COMMIT completes, the data has
been written to permanent storage.
If a database crash occurs, any transactions
will be COMMITed or not; no half-done
transactions.
No transaction can (directly) see another
transaction in progress.
In PostgreSQL
Everything runs inside of a transaction.
If no explicit transaction, each statement is
wrapped in one for you.
This has certain consequences for
database-modifying functions.
Everything that modifies the database is
transactional, even schema changes.
A brief warning
Many resources are held until the end of a
transaction.
Temporary tables, working memory,
locks, etc.
Keep transactions brief and to the point.
Be aware of IDLE IN TRANSACTION
sessions.
Transaction would be easy
if databases were single user.
Theyre not.
Thank goodness.
So, how do we handle concurrency control
when two sessions are trying to use the
same data?
The Problem.
Process 1 begins a transaction.
Process 2 begins a transaction.
Process 1 updates a tuple.
Process 2 reads that tuple.
What happens?
Bad Things.
Process 2 cant get the new version of the
tuple (ACID [generally] prohibits dirty
reads).
But where does it get the old version of
the tuple from?
Memory? Disk? Special roll-back area?
What if we touch 250,000,000 rows?
Some Approaches.
Lock the whole database.
Lock the whole table.
Lock that particular tuple.
Reconstruct the old state from a rollback
area.
None of these are particularly satisfactory.
Multi-Version Concurrency Control.
Create multiple versions of the database.
Each transaction sees its own version.
We call these snapshots in PostgreSQL.
Each snapshot is a first-class member of the
database.
There is no privileged real snapshot.
The Implications.
Readers do not block readers.
Readers do not block writers.
Writers do not block readers.
Writers only block writers to the same
tuple.
Snapshots.
Each transaction maintains its own
snapshot of the database.
This snapshot is created when a statement
or transaction starts (depending on the
transaction isolation mode).
The client only sees the changes in its own
snapshot.
Nothings Perfect.
PostgreSQL will not allow two snapshots
to fork the database.
If this happens, it resolves the conflict with
locking or with an error, depending on the
isolation mode.
Example: Two separate clients attempt to
update the same tuple.
Isolation Modes.
PostgreSQL supports:
READ COMMITTED The default.
REPEATABLE READ
SERIALIZABLE
It does not support:
READ UNCOMMITTED (dirty read)
What is a snapshot?
Logically, it is the set of all transactions that
have committed at a particular point in
time.
You can even manipulate snapshots (save
them, load them).
Snapshots are integral to how MVCC
works in PostgreSQL.
When does a snapshot begin?
In READ COMMITTED, each statement
starts its own snapshot.
Thus, it sees anything that has committed
since the last statement.
If it attempts to update a tuple another
transaction has touched, it blocks until that
transaction commits.
Higher isolation modes.
REPEATABLE READ and SERIALIZABLE
take the snapshot when the transaction
begins.
Snapshot lasts until the end.
An attempt to modify a tuple another
transaction has changed blocks
and returns an error if that
transaction commits.
Wait, what?
PostgreSQL attempts to maintain an
illusion of a perfect snapshot.
But if it cant, it throws an error.
The application then can retry the
transaction against the new, updated
snapshot.
SERIALIZABLE
Not every conflict can be detected at the
single tuple-level.
INSERTing calculated values.
SERIALIZABLE detects these using
predicate locking.
Requires some extra overhead, but
remarkably efficient.
MVCC consequences.
Deleted tuples are not (usually)
immediately freed.
Tuples on disk might not be available to
be readily checked.
This results in dead tuples in the database.
Which means:VACUUM!
VACUUM
VACUUMs primary job is to scavenge
tuples that are no longer visible to any
transaction.
They are returned to the free space for
reuse.
autovacuum generally handles this problem
for you without intervention.
ANALYZE
The planner requires statistics on each
table to make good guesses for how to
execute queries.
ANALYZE collects these statistics.
Done as part of VACUUM.
Always do it after major database changes
especially a restore from a backup.
Vacuums not working.
It probably is.
The database generally stabilize at 20% to
50% bloat. Thats acceptable.
If you see autovacuum workers running,
thats generally not a problem.
No, really, VACUUMs not working!
Long-running transactions, or idle-intransaction sessions?
Manual table locking?
Very high write-rate tables?
Many, many tables (10,000+)?
Unclogging the VACUUM.
Reduce the autovacuum sleep time.
Increase the number of autovacuum
workers.
Do low period manual VACUUMs.
Fix IIT sessions, long transactions, manual
locking.
Schema Design.
Whats Normal?
Normalization is important.
But dont obsess.
It flows naturally from proper separation of
data.
Pick Entities.
An entity is the top-level logical object in
your data model.
Customer, Order, InventoryItem.
Flow down from there to subsidiary items.
Make sure that no entity-level information
gets pushed into the subsidiary items.
Pick a naming scheme and stick with it.
Are tables plural or singular?
DB people tend to like plural, ORMs tend
to like singular.
Are field names CamelCase, lower_case, or
what?
Dont Repeat Yourself.
Denormalization generally means
including data that could be derived from
other sources.
Copied.
Calculated.
Calculated denormalization can sometimes
be useful; copied almost never.
Joins are Good.
PostgreSQL executes joins very efficiently.
Dont be afraid of them.
Especially dont worry about large tables
joining small tables.
PostgreSQL will almost always do the
right thing.
Use the Typing System.
PostgreSQL has a very rich set of types.
Use them!
If somethings a numeric, dont store it as a
string.
Use domains to create custom types.
No Polymorphic Fields.
Avoid fields whose interpretation is
dependent on another field.
Avoid fields which use strings to store
multiple types.
Keep each field well-defined as to what
data goes into it.
Constraints.
Use them. Theyre cheap and fast.
Constraints on single columns.
Constraints on multiple columns.
Exclusion constraints for constraints across
multiple rows.
Pick a naming scheme and stick with it.
Are tables plural or singular?
DB people tend to like plural, ORMs tend
to like singular.
Are field names CamelCase, lower_case, or
what?
Avoid Entity-Attribute-Value Schemas.
Each field should mean one thing, and one
thing only.
EAV schemas are nightmares to join and
report on.
They can also result in enormous database
bloat.
Key Selection.
SERIAL is convenient and straight-forward,
but
What if you have to merge two tables?
Use natural keys in preference to synthetic
keys if you possibly can.
Consider UUIDs instead of serials as
synthetic keys.
Dont Have Thing Tables.
OO programmers sometimes like to have
table hierarchies.
These tend to result in big base tables that
have common attributes factored out.
It looks normalized
but its really a pain in the neck.
Fast / Slow
If a table has a frequently-updated section
and a slowly-updated section, consider
splitting the table.
Do a 1:1 relationship between the two.
Keeps foreign key locking under control.
Arrays.
First-class type in PostgreSQL.
Can be searched, indexed, etc.
Often a good substitute to a subsidiary
table.
Often a great substitute to a big many-tomany table.
hstore
Much, much better than an EAV schema.
Great for optional, variable attributes.
Can be indexed, searched, etc.
But dont use it as a replacement for
schema modification!
JSON
First-class, in-core type.
Not quite as many search / indexing
operator as hstore
But its getting there.
Coming in 9.4: jsonb!
Indexing on Big Types.
PostgreSQL makes it work.
But it can be very inefficient.
Consider indexing on an expression of the
data:
Like the first 32 / last 16 characters of a
text string.
NULL
NULL is a total pain in the neck.
Sometimes, you have to deal with NULL,
but:
Only use it to mean missing value.
Never, ever have it as a meaningful value in
a key field.
WHERE NOT IN (SELECT ...)
Very Large Objects
Lets say 1MB or more.
Store them in files, store metadata in the
database.
The database API is not designed for
passing large objects around.
Many-to-Many Tables
These can get extremely large.
Consider replacing with array fields.
Either one way, or both directions.
Can use a trigger to maintain integrity.
Much smaller and more efficient.
Depends, of course, on usage model.
Character Encoding.
Use UTF-8.
Just. Do. It.
There is no compelling reason to use any
other character encoding.
One edge case: the bottleneck is sorting
text strings. This is very, very rare.
Time Representation.
Always use TIMESTAMPTZ.
TIMESTAMP is a bad idea.
TIMESTAMPTZ is timestamp, converted to
UTC.
TIMESTAMP is timestamp, at some time
zone but we dont know which one, hope
you do.
Indexing
Test your database
knowledge!
What does the SQL standard require for indexes?
Trick Question!
It doesnt.
The database should work identically
whether or not you have indexes.
Of course, identically in this case does
not mean just as fast.
No real-life database can work properly
without indexes.
PostgreSQL Index Types.
B-Tree.
Hash.
GiST.
SP-GiST.
GIN.
B-Tree Indexes.
The standard PostgreSQL index is a B-tree.
Provides O(log N) access to leaf notes.
Provides total ordering.
Operates on scalar values that implement
standard comparison operators.
B-Tree Index Types.
Single column.
Multiple column (composite).
Expression (functional) indexes.
Single Column B-Trees
The simplest index type.
Can be used to optimize searches on <,
<=, =, >=, >.
Can be used to retrieve rows in sorted
order on that column.
When to create?
If a query uses that column, and
uses one of the comparison
operators.
and selects <10-15% of the rows.
and is run frequently.
the index will likely be helpful.
Indexes and JOINs
Indexes can accelerate JOINs considerably.
But the usual rules apply.
Generally, they help the most when
indexing the key on the larger table and
that results in high selectivity against the
smaller table.
Indexes and Aggregates.
Some GROUP BY and related operations
can benefit from an index.
Often only in the presence of a HAVING
clause, though.
If it has to scan the whole index, it might as
well scan the whole table.
Mandatory indexes.
Constraints must have indexes to enforce
them.
Just accept those.
Ascending vs Descending?
By default, B-trees index in ascending order.
Descending indexes are much faster in
retrieving tuples in descending order.
So, if the primary function is descending
sortation, use that.
Otherwise, just use ascending order.
Composite Indexes.
A single index can have multiple columns.
The columns must be used left-to-right.
An index on (A, B, C) does not help a
query on just C.
But it does on (A, B).
Expression Indexes.
Indexes on an expression.
PostgreSQL can recognize when you are
querying on that expression and use the
index.
Can be expensive to create, but very fast to
execute.
Make sure PostgreSQL is really using it!
Partial Indexes.
An index does not have to contain all of
the rows of the table.
The WHEN clauses boolean predicate
limits the size of the index.
This can be a huge performance
improvement for queries that match the
predicate, all or in part.
Indexes and MVCC
The full key value is copied into the index.
Every version of the tuple on the disk
appears in the index.
Thus, PostgreSQL needs to check whether
a retrieved tuple is live.
This means indexes can bloat as dead
tuples pile up.
GiST Indexes.
GiST is not a single index type, but an index
framework.
It can be used to create B-tree-style
indexes.
It can also be used to create other index
types, like bounding-box and geometric
queries.
GiST Index Usage.
Non-total-ordered types generally require
a GIST index.
Each types index implementation decides
what operators to support.
Inclusion, membership, intersection
Some GiST indexes do provide ordering.
KNN indexes, for example.
GIN
Generalized Inverted Index.
Maps index items (words, dict keys) to
rows whose field contains those.
Core PostgreSQL use: Full text search
indexes.
Maps tokenized words to the rows
containing those words.
GIN implementation
A B-tree of B-trees.
Tokens organized into B-trees.
Row pointers also organized into B-trees.
On-disk footprint can be quite large.
Why isnt it using my indexes?
The most common complaint.
First, get the EXPLAIN ANALYZE output of
the query.
Sometimes, it is using the index, and its just
slow anyway!
Bad Selectivity.
If PostgreSQL thinks that the index scan
will return a large percentage of the table, it
will do a seq scan instead.
Generally, its right to think this.
If its wrong, and the query is very
selective, try re-running ANALYZE.
ANALYZE didnt help.
Try running the query with:
SET enable_seqscan = off;
See how long it takes to use the index
then.
PostgreSQL might be right.
Hey, it didnt use the index even then!
Index Prohibitorum
This means PostgreSQL thinks that index
doesnt apply to this query.
Query mis-written? Index invalid?
Confusing expression index?
Try doing a very simple query on just that
field, and build up.
PostgreSQL is right, but wrong.
In fact, using the index is faster even though
PostgreSQL thinks it is not.
Try lowering random_page_cost.
Consider changing the default statistics
target for that field.
PostgreSQL, Your Query Plan Sucks.
Bitmap Heap Scan on mytable (cost=12.04..1632.35 rows=425
width=321)
Recheck Cond: (p_id = 543094)
-> Bitmap Index Scan on idx_mytable_p_id
(cost=0.00..11.93 rows=425 width=0)
Index Cond: (p_id = 543094)
What does this mean?
First, PostgreSQL scans the index and
builds a bitmap of pages (not tuples!) that
contain candidate results.
Then, it scans the heap (the actual
database), retrieving those pages.
And then rechecks the condition against
the tuples on that page.
That makes no sense whatsoever.
PostgreSQL does this when the number of
tuples to be retrieved is large.
It can avoid doing lots of random access to
the disk.
Pure Index Scan.
Index Scan using testi on test
width=4)
Index Cond: (whatever = 5)
(2 rows)
(cost=0.00..8.27 rows=1
Index Creation.
Two ways of creating an index:
CREATE INDEX
CREATE INDEX CONCURRENTLY
CREATE INDEX
Does a single scan of the table, building the
index.
Uses maintenance_work_mem to do the
creation.
Keeps an exclusive lock on the table while
the index build is going on.
CREATE INDEX CONCURRENTLY
Does two passes over the table:
Builds the index.
Validates the index.
If the validation fails, the index is marked as
invalid and wont be used.
Drop it, run again.
REINDEX
Rebuilds an existing index from scratch.
Takes an exclusive lock on the table.
Generally no need to do this unless an
index has gotten badly bloated.
Index Bloat.
Over time, B-tree indexes can become
bloated.
Sparse deletions from within the index
range are the usual cause.
http://pgsql.tapoueh.org/site/html/news/
20080131.bloat.html
Generally, dont worry about it.
Index Usage.
pg_stat_user_indexes
Reports the number of times an index is
used.
If non-constraint indexes are not being
used, drop them.
Indexes are very expensive to maintain.
Debugging
This query is slow.
EXPLAIN or EXPLAIN ANALYZE
The output is somewhat cryptic.
Lets look at an example from the bottom
up.
http://explain.depesz.com/
select COUNT(DISTINCT "ecommerce_order"."id") FROM
"ecommerce_order" LEFT OUTER JOIN "ecommerce_solditem" ON
("ecommerce_order"."id" = "ecommerce_solditem"."order_id") WHERE
("ecommerce_order"."subscriber_id" = 396760 AND
("ecommerce_solditem"."status" = 1 AND
("ecommerce_solditem"."user_access_denied" IS NULL OR
"ecommerce_solditem"."user_access_denied" = false ) AND
"ecommerce_order"."status" IN (3,9,12,16,14)));
---------------------------------------------------------------------Aggregate (cost=2550.42..2550.43 rows=1 width=4)
-> Nested Loop (cost=0.00..2550.41 rows=3 width=4)
-> Index Scan using ecommerce_order_subscriber_id
on ecommerce_order (cost=0.00..132.88 rows=16 width=4)
Index Cond: (subscriber_id = 396760)
Filter: (status = ANY ('{3,9,12,16,14}'::integer[]))
-> Index Scan using ecommerce_solditem_order_id
on ecommerce_solditem (cost=0.00..150.86
rows=19 width=4)
Index Cond: (ecommerce_solditem.order_id =
ecommerce_order.id)
Filter: (((ecommerce_solditem.user_access_denied
IS NULL) OR
(NOT ecommerce_solditem.user_access_denied))
AND (ecommerce_solditem.status = 1))
Query Analysis.
Read the execution plan from the bottom
up.
Look for nodes that are processing a lot of
data
especially if the data set is being
reduced considerably on the way up.
ANALYZE
The planner requires good statistics to
create these plans.
ANALYZE collects them.
If the statistics are bad, the plans will be,
too.
---------------------------------------------------------------------Aggregate (cost=48353.52..48353.53 rows=1 width=4)
-> Nested Loop (cost=0.00..48353.52 rows=1 width=4)
-> Seq Scan on ecommerce_solditem
(cost=0.00..38883.38 rows=868 width=4)
Filter: (((user_access_denied IS NULL) OR
(NOT user_access_denied)) AND (status = 1))
-> Index Scan using ecommerce_order_pkey on
ecommerce_order (cost=0.00..10.90 rows=1 width=4)
Index Cond: (id = ecommerce_solditem.order_id)
Filter: ((subscriber_id = 396760) AND
(status = ANY ('{3,9,12,16,14}'::integer[])))
---------------------------------------------------------------------Aggregate (cost=2550.42..2550.43 rows=1 width=4)
-> Nested Loop (cost=0.00..2550.41 rows=3 width=4)
-> Index Scan using ecommerce_order_subscriber_id
on ecommerce_order (cost=0.00..132.88 rows=16 width=4)
Index Cond: (subscriber_id = 396760)
Filter: (status = ANY ('{3,9,12,16,14}'::integer[]))
-> Index Scan using ecommerce_solditem_order_id
on ecommerce_solditem (cost=0.00..150.86
rows=19 width=4)
Index Cond: (ecommerce_solditem.order_id =
ecommerce_order.id)
Filter: (((ecommerce_solditem.user_access_denied
IS NULL) OR
(NOT ecommerce_solditem.user_access_denied))
AND (ecommerce_solditem.status = 1))
Planner Statistics
Collected as histograms on a per-column
basis.
100 buckets by default.
Not restored from backup!
Not automatically updated on major
database updates!
Cost.
Measured in arbitrary units (traditionally
have been disk fetches).
First number is the startup cost for the
first tuple, second is the total cost.
Comparable with other plans using the
same planner configuration parameters.
Costs are inclusive of subnodes.
Actual Time.
In milliseconds.
Wall-clock time, not only query execution
time.
Also presents startup time, total time.
Also inclusive of subnodes.
Rows.
Estimated and actual rows emitted by each
planner node.
Not the number processed; that could be
larger, and is reflected in cost.
A large mismatch is one of the first places
to look for query problems.
Loops.
Number of times a subplan was executed
by its parent.
In this case, actual times are averages, not
totals.
Things that are bad.
JOINs between two very large tables.
Very difficult to execute efficiently unless
the sides can be reduced by a predicate.
CROSS JOINs
These can be created by accident!
Sequential scans on large tables.
SELECT COUNT(*)
Always results in a full table scan in
PostgreSQL.
So dont do that.
OFFSET / LIMIT
Everyones favorite way of implementing
pagination.
OK for low OFFSET values
but comes apart fast for higher ones.
GoogleBot Is Relentless.
Precalculate, use other keys.
The database is slow.
Whats going on?
pg_stat_activity
tail -f the logs.
Too much I/O? iostat 5
The database isnt responding.
Make sure its up!
Can you connect with psql?
pg_stat_activity
pg_locks
Python Particulars
Python 2? psycopg2
Overall, the best library for accessing
PostgreSQL directly from Python.
Hard to justify using anything else.
Very feature-rich, very Pythonic (such as
DB API 2 is Pythonic).
psycopg2 notes.
The result set of a query is loaded into
client memory when the query
completes
regardless of the size of the result set!
If you want to scroll through the results,
used named cursors.
py-postgresql
Python 3.x driver.
Pure Python, so can run under interpreters
that require it.
Django Notes.
If you are running on 1.6+, always use the
@atomic decorator.
Cluster write operations into small
transactions, leave read operations outside.
Do all your writes at the very end of the
view function.
Django + Replication
Multi-database works very nicely with hot
standby.
Point the writes at the primary, the reads at
the secondary.
Django 1.5 or earlier.
Use the @xact decorator and style.
https://github.com/xof/xact
Sloppy transaction management can cause
the dreaded Django idle-in-transaction
problem.
Go South.
Use South in Django for migration
management.
Create manual migrations for schema
changes that Django cant specify.
Specialized constraints, indexes, etc.
Special Situations.
Minor version upgrade.
Do this promptly!
Only requires installing new binaries.
If using packages, often as easy as just an
apt-get / yum upgrade.
Very small amount of downtime.
Major version upgrade.
Requires a bit more planning.
pg_upgrade is now reliable.
Trigger-based replication is another option
for zero downtime.
A full pg_dump / pg_restore is always
safest, if practical.
Always read the release notes!
Dont get caught!
Major versions are EOLd after 5 years.
Always have a plan for how you are going
to move between major versions.
All parts of a replication set must be
upgraded at once (for major versions).
Bulk loading data.
Use COPY, not INSERT.
psycopg2 has a very nice COPY interface.
COPY does full integrity checking and
trigger processing.
Do a VACUUM afterwards.
Very high insert rates.
Reduce shared buffers by 25%-75%.
Reduce checkpoint timeouts to 3min or
less.
Make sure to do enough ANALYZEs to
keep the statistics up to date, manual if
required.
AWS
Generally, works like any other system.
Remember that instances can disappear and
come back up without instance storage.
Always have a good backup / replication
implementation on AWS!
PIOPS are useful (but pricey) if you are
using EBS.
Larger-Scale AWS Deployments
Script everything: Instance creation,
PostgreSQL setup, etc.
Put everything inside a VPC.
Scale up and down as required to meet
load.
AWS is a very expensive equipment
rental service.
PostgreSQL RDS
Overall, not a bad product.
BIG plus: Automatic failover.
BIG minus: No reading from the secondary.
Other minuses: Expensive, fixed (although
large) set of extensions.
Not a bad place to start with PostgreSQL.
Sharding.
Eventually, you will run out of write
capacity on your master.
Then what?
Community PostgreSQL doesnt have an
integrated multi-master solution.
But there are options!
Postgres-XC
Open-source fork of PostgreSQL.
Intended for dedicated hardware in a single
rack.
Node failure is still a challenge.
Somewhat experimental, but shows great
promise.
Bucardo
Has multi-master write capability.
Handles burst-writes effectively.
Not great for sustained writes, since the
writes ultimately have to end up on all
machines.
Custom Sharding.
Distribute data across multiple machines in
a way that the application can find it.
Can shard on an arbitrary value (user ID),
or something less abstract (region).
Application is responsible for routing to the
right database node.
http://instagram-engineering.tumblr.com/
post/10853187575/sharding-ids-atinstagram
Pooling, etc.
Why pooling?
Opening a connection to PostgreSQL is
expensive.
It can easily be longer than the actual query
time.
Above 200-300 connections, use a pooler.
pgbouncer
Developed by Skype.
Easy to install.
Very fast, can handle 1000s of connections.
Does not to failover, load-balancing.
Use HAProxy or similar.
pgPool II
Does query analysis.
Can route queries between master and
secondary in replication pairs.
Can do load balancing, failover, and
secondary promotion.
Higher overhead, more complex to
configure.
Tools
Monitor, monitor, monitor.
Use Nagios / Ganglia to monitor:
Disk space at minimum.
CPU usage
Memory usage
Replication lag.
check_postgres.pl (bucardo.org)
Graphical clients
pgAdmin III
Comprehensive, open-source.
Navicat
Commercial product, not PostgreSQLspecific.
Log Analysis
pgbadger
The only choice now for monitoring text
logs.
pg_stat_statements
Maintains a buffer of data on statements
executed, within PostgreSQL.
Questions?
thebuild.com / @xof / pgexperts.com
Thank you!
http://tinyurl.com/pycon2014survey
thebuild.com / @xof / pgexperts.com
