postgres-xc: symmetric postgresql cluster

Postgres-XCSymmetric PostgreSQL Cluster

Pavan DeolaseeJuly 29, 2013

PostgreSQL Global Development GroupPostgres-XC Global Development Group

11/21/13 Postgres-XC 2

Who am I ?

Pavan Deolasee� Http://www.linkedin.com/in/pavandeolasee

Contributor to PostgreSQL and Postgres-XC Global Development� Best known for development of Heap-Only-Tuple (HOT) feature of PostgreSQL

8.3 release

� Contributed several other enhancements to PostgreSQL and derivatives

� Associated with Postgres-XC since its inception

� Contributed to Postgres-XC's architectural design and implementation of several features

Previously worked for EnterpriseDB and Symantec/Veritas

http://www.linkedin.com/in/pavandeolasee


Agenda

PostgreSQL

History of Postgres-XC

Architecture Overview

Data Distribution Strategies

Scalability and Performance Results

SPOF Analysis

Development Practices, Project Status and Roadmap

Support

PostgreSQL


PostgreSQL Recap

PostgreSQL� is world's most advanced open source database

� is very stable

� is fully compliant with ANSI SQL

� supports foreign key, check constraints

� supports various kinds of indexes

� supports inheritance

� is fully extensible (data types, procedural languages etc)

� supports ACID transactions

� will recover your database in case of server failure


PostgreSQL Recap

PostgreSQL� uses write-ahead-logs for durability and recover your database in case of server

failure

� built-in log based streaming synchronous/asynchronous replication

� file system level backups and archive recovery

� point-in-time recovery

� hot standby

� binary upgrades

� full-text search

� and many more


Performance and Scalability

While PostgreSQL was becoming feature rich, scalability was still a concern, especially on high end machines.

Significant work in last few releases to address performance and scalability issues

� Heap-only-tuples (HOT) in 8.3 for reuse of dead space

� Virtual transaction identifiers for read-only transactions

� Spread checkpoints to reduce spikes in disk IO

� Hot standby for read scalability

� Visibility maps, Index-only-scans

� Improved spin-lock implementations

� Linear read scalability for multi-core machines (64 cores)

� Unlogged tables



Scale up is a continuous process� Every major release comes up with new improvements

� Reduced contention for critical sections is important for multiprocess scalability

64 and 128 core machines are becoming common Several hundred concurrent connections is a norm

� Efficient and full use of resources on high end machines

Large amounts of RAM

While scale up is nice and very important, it may not be able to keep pace with the volume and growth of data

Streaming replication and Hot standby allows read scalability, but each server should still be capable of handling full data

Partitioning using inheritance and constraint exclusion



When data grows rapidly, it may not be possible or affordable to upgrade hardware quickly

Scale out is a very attractive solution for BIG data use cases

PostgreSQL had some support as contrib module� DBlink

Foreign Data Wrappers (FDW)� Create foreign servers and foreign tables

� Natively query those tables from PostgreSQL server

� FDWs exist for Postgres, MySQL, File, MongoDB etc

� Though attractive, still in very early stage and may not scale with complex queries

History of Postgres-XC


Project Kick-off

2009 - NTT Data Intellilink Corporation, Japan and EnterpriseDB (my previous employer) decide to collaborate for developing a clustering solution for PostgreSQL

� NTT had previously internally tried to build another Postgres cluster (RITA-DB), but could not take it beyond prototype stage. Lessons learnt

� EnterpriseDB had previously acquired GridSQL, a scale-out solution. But it had its own issues including performance problems

� No other existing clustering solution was good enough

Postgres-R Slony-II Pgpool-II GridSQL Log Shipping Standby pl/proxy


Project Mandate

Provide a viable open source, free alternative to a popular shared disk cluster from a very reputable commercial vendor. Yes, Oracle RAC

Provide write-scalability� At least 3.5 times scalability with 5 servers in a write intensive workload

Provide read-scalability

� Efficient utilization of cluster nodes

� Parallel execution


Project Mandate (Cont)

Provide global consistency� Data written from one node must be available from other nodes immediately and

in a consistent manner

Provide ACID properties of transaction

Provide seamless API so that the client applications need no change or very minor change

� Same SQL

� Same libpq, JDBC, ODBC APIs

Must run and scale on commodity hardware

� Expensive shared storage

� Expensive network interconnects


Initial Team

Koichi Suzuki� Chief architect and leader of the project, NTT Data Intellilink

Mason Sharp� StormDB (previously at EnterpriseDB)

Pavan Deolasee (me)

� (previously at EnterpriseDB)

Many more from NTT, EnterpriseDB and open source community has joined and contributed since then

(Part of the content is attributed to this team and the global development team)

Postgres-XC Architecture


So What is Postgres-XC ?

Shared nothing architecture

Symmetric PostgreSQL cluster� No master/slave replication

� No read-only clusters

� Every node can issue both read/write

� Every node provides single consistent database view

� Transparent transaction management

Not just a replication� Each table can be replicated/distributed by sharding

� Parallel transaction/query execution

So both read/write scalability


PostgreSQL Streaming Replication

Master-slave architecture

Write-Ahead-Log (WAL) based replication

Synchronous, asynchronous, cascading replication supported

Hot Standby� Can answer read-only queries

� Standby may give a stale view of the database since the WAL application happens asynchronously

� Do not support table level replication


Postgres-XC: Symmetric Cluster

Global TransactionManager (GTM)

Postgres-XC Servers


Postgres-XC Key Points

Postgres-XC is symmetric, shared nothing, scalable PostgreSQL cluster

Write Scalability

Read Scalability

Global Consistency


Postgres-XC Major Components

Applications

Coordinators

Datanodes

GTM

SQL/libpq Interface

SQL / libpq / ODBC / JDBC Interface


PostgreSQL MVCC

PostgreSQL uses Multi-version concurrency control (MVCC) for guaranteeing Isolation and Consistency (“CI” of ACID)

MVCC increases concurrency � readers don't wait on other readers/writers

� writers don't wait on readers

Every write transaction is assigned an XID (32-bit identifier)

Every row that gets inserted/deleted/updated gets a stamp of the transaction that did the operation

A MVCC Snapshot is used to decide if a given row or a version of a given row is visible to a transaction or not


PostgreSQL MVCC

Two important concepts

� Transaction Identifiers (XIDs)

� MVCC Snapshots

A consistent view of the database can be obtained by ensuring that these two values are managed properly


Global Transaction Manager (GTM)

The only central component in the XC cluster

Provides globally consistent and visible transaction identifiers (GXID) to transactions running across the cluster

Provides global snapshots to run queries across the cluster so that they see a consistent view irrespective of how many different nodes the query need to fetch data from

GTM is not a database itself, though it keeps some simple state information

Its a separate binary supplied with Postgres-XC source

Runs as a independent server process


Global Transaction Manager (GTM)

There can only be ONE GTM per Postgres-XC cluster, though additional GTM standby can be configured (more later)

Comes with a GTM-proxy (more later) for request pooling and scalability

Multi-threaded process (one thread per connection)

Recommended to run GTM on a separate physical machine for performance and scalability

Helps in a few other side issues such as to manage sequences


Coordinator

Coordinator is the brain of Postgres-XC cluster

Applications connect to the coordinators and issue SQL queries (they never talk to datanodes directly)

� Parse, plan and execute queries

� Knows about table distribution properties and takes advantage of the same for query planning

� Pushes down as much computation as possible to the datanodes

� Pushes down transaction identifiers and snapshots to other nodes

� Manages implicit two-phase transactions


Coordinator

Non-critical component of the Postgres-XC cluster

There could be (and most likely there will be) more than one coordinator in the cluster

Applications can connect to any coordinator and issue read/write queries

� No master coordinator

Modified version of PostgreSQL� Remember PostgreSQL comes with PostgreSQL license


Datanode

The real guy managing all the data

Accepts requests from coordinators and sends responses back� SQL/DML queries

� DDL queries

� Maintenance commands (VACUUM, CHECKPOINT, CLUSTER etc)

Modified PostgreSQL engine but with very minimal changes

� Transaction management, MVCC

� Sequences

Critical component of the cluster� Failure will lead to DML/DDL queries failing unless the coordinator can decide to

bypass the datanode because its not required for the given query

Data Distribution


Distribution Strategies

Distributed tables� Each row exists only on a single datanode

� Distribution strategies

HASH MODULO ROUND ROBIN User defined functions (Future) Range (Future) Values (Future)

Replicated tables� Each row of the table is stored on all datanodes where the table is replicated


Distributed Tables

Write to a single row is applied only on the node where the row resides

� Multiple rows can be written in parallel

� Scanning rows spanning across the nodes (e.g. table scans) can hamper performance

Point reads and writes based on the distribution column value show good performance

Datanode where the operation happens can be identified by the distribution column value in the query


Replicated Tables

Statement level replication

� Each write needs to be replicated

� Writes are costly

Read can happen on any node (where table is replicated)� Reads from different coordinators can be routed to different nodes

Useful for relatively static tables, with high read load


Distribution Samples – Horizontal Scalability

CREATE TABLE emp (empno integer, ename text) DISTRIBUTE BY HASH(empno);

CREATE TABLE emp (empno integer, ename text) DISTRIBUTE BY MODULO(empno);

CREATE TABLE emp (empno integer, ename text) DISTRIBUTE BY ROUND ROBIN;

CREATE TABLE emp (empno integer, ename text);

Default is to distribute by HASH if the datatype is supported

� first column of primary key or� first column of foreign key or� first column of the table� If everything else fails, distribute by ROUND ROBIN


Distribution Samples (Cont)

CREATE TABLE emp (empno integer, ename text) DISTRIBUTE BY HASH(empno)

TO NODE (N1, N2);

Distribute to all nodes or a set of nodes, either explicitly named or grouped together as named group

Default is to distribute to all the nodes configured


Distributed Query Processing

Coordinator

� Accepts queries and plans them

� Finds the right data-nodes from where to fetch the data

� Frames queries to be sent to these data-nodes

� Gathers data from data-nodes

� Processes it to get the desired result

Datanode� Executes queries from coordinator like PostgreSQL

� Has same capabilities as PostgreSQL


Distributed Query Processing

Query shipping

Coordinator tries to delegate maximum query processing to data-nodes

� Indexes are located on datanodes

� Materialization of huge results is avoided in case of sorting, aggregation, grouping, JOINs etc.

� Limit, Offset clauses are shipped to the datanodes

� ORDER BY, GROUP BY clauses are shipped to the datanodes

� Coordinator is freed to handle large number of connections

� Joins are shipped too

Distributing data wisely helps coordinator to delegate maximum query processing and improve performance

Scalability by Data Distribution


Write Scalability

Tables are distributed/sharded – usually HASH of the column value

� Horizontal partitioning

Inserts are directed to the datanode based on the value of the distributed column

� Integrated in the query executor

� Not trigger or rule based

� Supports COPY (bulk load mechanism of PostgreSQL)

Updates and Deletes also directed to the datanodes� Coordinator is completely bypassed if the query does not have any local/cross

dependency

� Highly scalable in many cases

Some queries must go through the coordinator because of dependencies


Write Scalability (Cont)

CREATE TABLE test (a int, b text) DISTRIBUTE BY HASH(a);

INSERT INTO test SELECT generate_series(1,100), 'foo';

EXPLAIN VERBOSE UPDATE test SET b = 'bar' WHERE a > 1;

QUERY PLAN

----------------------------------------------------------------------------

Data Node Scan on "__REMOTE_FQS_QUERY__" (cost=0.00..0.00 rows=0 width=0)

Output: ('bar'::text), ('bar'::text), test.ctid, xc_node_id

Node/s: node_dn1, node_dn2

Remote query: UPDATE public.test SET b = 'bar'::text WHERE (a > 1)

(4 rows)


Write Scalability (Cont)

EXPLAIN VERBOSE DELETE FROM test WHERE a = 100;

QUERY PLAN

----------------------------------------------------------------------------

Data Node Scan on "__REMOTE_FQS_QUERY__" (cost=0.00..0.00 rows=0 width=0)

Output: test.a, test.ctid, xc_node_id

Node/s: node_dn2

Remote query: DELETE FROM public.test WHERE (a = 100)

(4 rows)

Queries are started in parallel on multiple nodes thus improving scalability of write operations


Read Scalability

Significant planner enhancements to improve read scalability� Many cool features already in

� Continuous development for improving it further

Query is analyzed and shipped to a datanode as it is if there are no dependencies on the coordinator or other datanodes

� Fast Query Shipping

If entire query can not be shipped, parts of the query are shipped� Offload as much processing as possible to the datanodes

� Do the final processing of data on the coordinator

� Reduce as much data movement from datanodes to the coordinator


Read Scalability (Cont)

Parallel query execution

� Query is stared in parallel on all involved datanodes

� Significantly improves performance by utilizing capacities across the cluster

Qualifications push-down� WHERE clauses in a SELECT query are pushed down to the datanodes to

reduce the selectivity of the result set

� Significantly reduces the amount of data fetched at the coordinator for final execution

Remote expressions push-down� Push function and expression evaluation to the datanodes



Aggregates� Enhanced PostgreSQL's two step aggregation to three step aggregation

� Collect summary at each datanode and then do the final aggregation at the coordinator

An example, avg(integer)

Datanode D1 - 1, 2, 5, 7

Datanode D2 - 3, 4, 6, 8, 9

Datanode D1 sends: sum(integer), count => 15, 4

Datanode D2 sends: sum(integer), count => 30, 5

Coordinator computes the final avg as (15 + 30) / (4 + 5) = 5



Join Push-down

� Complete or part of complex joins can be pushed down to the datanodes

Complex analysis of the tables involved in the join and their distribution strategy is required to arrive at a conclusion regarding what can be pushed and what must not be pushed

Significant performance and scalability boost

Order By� Result sets are sorted on individual datanodes and fetched on the coordinator

� Coordinator does a final merge step of the sorted sets

Limit/Offset


Thumb Rules - Read-Write Queries

High point reads (based on distribution column)

� Distributed or replicated

High read activities but no frequent writes� Better be replicated

High point writes� Better be distributed

High insert-load, but no frequent update/delete/read� Better be round-robin


Query Analysis

Find the relations/columns participating in equi-join conditions, WHERE clause etc.

� Distribute on those columns

Find columns participating in GROUP BY, DISTINCT clauses


Find columns/tables which are part of primary key and foreign key constraints

� Global constraints are not yet supported in XC



Example DBT1


Example DBT1

author, item� Less frequently written

� Frequently read from

� Author and item are frequently JOINed

Dimension tables� Hence replicated on all nodes


Example DBT1

customer, address, orders, order_line, cc_xacts

� Frequently written

hence distributed� Participate in JOINs amongst each other with customer_id as JOIN key

� point SELECTs based on customer_id

hence distributed by hash on customer_id so that JOINs are shippable� Participate in JOINs with item

Having item replicated helps pushing JOINs to datanodes


Example DBT1

Shopping_cart, shopping_cart_line

� Frequently written

Hence distributed� Point selects based on column shopping_cart_id

Hence distributed by hash on shopping_cart_id� JOINs with item

Having item replicated helps


Results – DBT1

Other Scalability Considerations


GTM Scalability

GTM can become a bottleneck

� serve XID for every new transaction in the cluster

� MVCC snapshot for transaction or every statement in a transaction (read-committed transactions)

� Snapshot size can increases proportional to the number of concurrent transactions in the system

GTM has been heavily optimized and runs a multi-threaded process for scalability

For even greater scalability, advised to run GTM Proxy on every coordinator


GTM Scalability – GTM Proxy

Runs as a proxy to GTM on other servers – preferably on the same server where coordinators run

� One proxy per coordinator server

Same API as GTM� Components can connect to a proxy or GTM directly

Pools multiple GET TXID requests into a single request

� e.g get 100 new identifiers in a single request/response

Pools multiple GET SNAPSHOT requests into a single request� Some manipulation of the received snapshot is required, but that's possible

Further pools requests of multiple types� Reduces load on the GTM significantly

� One proxy per server scales near linearly


Coordinator Scalability

Coordinator does not store table data

� IO scalability not an issue

Query parsing, planning and part or full execution� CPU utilization could be high

Add more coordinators if you need more computing power

Beware of connections explosion on the datanodes� Every additional coordinator will need more connections to datanode

� Built-in connection pooler in the coordinator can help a lot


Datanode Scalability

Datanode stores and fetches table data from the disks

In case of IO saturation� Add more datanodes and reconfigure existing tables

Easier said than done May need data redistribution (except for round-robin distribution) Table remains locked out (until concurrent data redistribution is added)

� Add more datanodes for new tables

Create new tables on the new datanode Existing tables are not impacted


Transaction Management

Postgres-XC uses two-phase commit protocol for distributed transaction management

� Every commit will make two round-trips to the datanode (PREPARE and COMMIT)

� Optimized so that single phase commit is used if only one node is involved in a transaction

� Optimized so that only write-operating datanodes/coordinators are counted for deciding if 2PC is required or not

If possible, design transactions so that they span a single or as few nodes as possible

SFOF Analysis and Failure Handling


Single Point of Failure

GTM

� Obviously SPOF

GTM-Proxy� No persistent data

� Just restart when fails

Coordinator

� Every coordinator is essentially a copy

� When fails, other coordinators work

Datanode� SPOF for sharded table


SPOF - GTM

GTM keeps crucial information about next transaction identifier, currently running transactions as well as sequence information

� Actual fate of each transaction is not permanently saved on the GTM though (at least not today)

GTM failure leads to complete failure of the cluster� any useful work would needs GTM interaction

GTM Standby for rescue


GTM Standby

GTM Standby is a working replica of the GTM

� Every operation on the GTM is replicated on the GTM, either via proxy or GTM

Create and configure standby� Only one standby can be configured currently

� Can be added to a running cluster without any disruption

� Can be promoted to be the new master

� Same executable as GTM, only configuration change


SPOF - Datanode

Almost all the techniques for PostgreSQL HA are available

� Streaming replication

� Shared disk remount

Coordinators may require reconfiguration� Coordinators should use the new datanode

� Coordinators should clean connections to failed datanode before reconfiguration

GTM� Reconnect to (new) local GTM proxy


SPOF – Coordinator

Coordinator only maintains catalog information

� Stable and mostly read-only

All coordinators are replica of each other� No master-slave concept. Every coordinator is equal

Datanode HA techniques can be applied� Streaming replication

� Shared disk remount

Applications can continue to use other coordinators while the failed coordinator is restored

Development Process and Status


Development Process

Project hosted on sourceforge

� http://postgres-xc.sourceforge.net/

Join developer/user mailing lists� [email protected]

� [email protected]

Test and report bugs

� [email protected]

Submit patches

http://postgres-xc.sourceforge.net/

mailto:[email protected]




Regression and Performance Tests

PostgreSQL's regression tests are enhanced

� New tests added to test Postgres-XC specific features

� Existing tests modified to adjust for difference because of cluster environment

Nightly regression tests are run on buildfarm (at least last I checked)

Regular performance tests are conducted by NTT Data to guard against performance regression

� Typically DBT1 tests

� Last I heard, DBT2 tests were being developed by NTT


Status

Release 1.1 is in beta stage (recommended)� Based on latest PostgreSQL 9.2 stable release

� Many new features, enhancements and bug fixes

Data redistribution Triggers Planner enhancements GTM standby Where current of Returning

Release 1.0 is the latest stable release� Based on PostgreSQL 9.1


Release Management

Yearly major release

� Merge with PostgreSQL's new release

� Additional XC features

Quarterly minor release� Bug fixes

� Security fixes

� Catch up with PostgreSQL's minor release


Licensing

Started with GPL license, but changed to PostgreSQL license couple of years back

� Anyone can take the source, modify and use without requiring to open source the modified code

� Very flexible licensing


Support

Very good support available from the core team on the mailing lists

� Helps if a self containing test case is included

Encouraged to fix bugs on your own and submit a patch

A committer will look at the patch and commit (with changes if required)

Volume of bug reports is still low and hence good turnaround time

A few commercial companies have started offering professional support

� Good for project in the long run

� Skills of support staff could be limited right now

Limitations and Roadmap


Current Limitations

Multi-column distribution is not supported

Global constraints not supported� Unique index on non-distribution column is not supported

Updating distribution column not supported

Statement level triggers not supported on the datanode side� Statement level triggers will fire at the coordinator


Current Limitations

No cluster wide deadlock detection

� Timeout based deadlock resolution

No cost estimation at coordinator (table statistics not available/used at the coordinator)

� ANALYZE command is propagated to the datanodes and the stats are updated on the datanodes

� Datanodes have up-to-date statistics. So they will plan the queries in the most efficient way

Node management needs much improvement� No built-in HA for replicated tables SELECT queries


Current Limitations

Truly serializable transactions not supported

� Read-committed and repeatable reads are supported

Transactions accessing TEMP objects can not be committed with 2PC

� PostgreSQL does not support PREPARing a transaction that has accessed temporary objects

� Set enforce_two_phase_commit = off to handle such transactions

Savepoints not supported

Exception blocks in plpgsql not supported (they need savepoint support)


Roadmap

Support for remaining SQL features (continuous process)

Planner enhancements so that more and complex queries can be pushed down to the datanodes

Table statistics at the coordinator

Concurrent redistribution of data after node additional/removal� Without exclusive lock on the table

Separate group in NTT is developing resource agents for corosync/pacemaker

� Plans not yet in public domain

Installers and further testing with DBT-2/DBT-5

Resources


Project Resources

Postgres-XC project home page

� http://postgres-xc.sourceforge.net/

� Project wiki and links to other presentations/white papers

Downloads and installation� Product RPMS available for download

Source code

� Source code from GIT repository

� Zipped tar balls

Other tools� pgxc_ctl utility to configure/manage/monitor the cluster

� A bunch of other contrib modules

http://postgres-xc.sourceforge.net/


Project Resources

Documentation� Extensive Postgres style documentation explaining XC-specific features

� Available online as well as distributed with source/rpms.

Mailing lists� Postgres-xc-general

� Postgres-xc-developers

� Postgres-xc-bugs

Write to me� [email protected]

Thank You

postgres-xc: symmetric postgresql cluster

Technology