Building and deploying large scale real time news system with my sql and distributed cache mysql_conf

Building and Deploying Large Scale
Real Time News System with
MySQL and Distributed Cache
Presented
to
MySQL
Conference

Apr.
13,
2011

Who am I?
Pag
e2

  Tao Cheng <tao.cheng@teamaol.com>, AOL Real
Time News (RTN).
  Worked on Mail and Browser clients in the ‘90 and
then moved to web backend servers since.
  Not an expert but am happy to share my experience
and brainstorm solutions.

Presentation for
[CLIENT]

Agenda

  AOL Real Time News (RTN): what it is?
  Requirements
  Technical solutions with focus on MySQL
  Deployment Topology
  Operational Monitoring
  Metrics Collection

Agenda

  Tips for query tuning and optimization
  Heuristic Query Optimization Algorithm
  Lessons learned
  Q & A

Real Time News : background
Pag
e5

AOL deployed its large scale Real Time News (RTN)
system in 2007.
This system ingests and processes news from 30,000
sources on every second around the clock. Today, its
data store, MySQL, has accumulated over several
billions of rows and terabytes of data.
However, news are delivered to end users in close to
real time fashion. This presentation shares how it is
done and the lessons learned.

Presentation for
AOLU Un-University

Brief Intro: sample features
Pag
e6

  Data presentation: return most recent news in
  flat view – most recent news about an entity. An entity could
be a person, a company, a sports team, etc.
  topic clusters – most recent news grouped by topics. A topic is
a group of news about an event, headline news, etc.
  News filtering by
  source types such as news, blogs, press releases, regional, etc.

  relevancy level (high, medium, low, etc) to the entities .

  Data Delivery: push (to subscribers) and pull
  Search by entities, categories (National, Sports,
Finance, etc), topics, document ID, etc.
Presentation for
[CLIENT]

Requirements for Phase I (2006)
Pag
e7

  Commodity hardware: 4 CPU, 16 GB MEM, 600 GB
disk space.
  Data ingestion rate = 250K docs/day; average
document size = 5 KB.
  Data retention period: 7 days to forever
  Est. data set size: (1.25 GB/day or 456 GB/year) +
space for indexes, schema change, and optimization.
  Response time: < 30 milli-second/query
  Throughputs: > 400 queries/sec/server
  Up time: 99.999%
Presentation for
[CLIENT]

Solutions: MySQL + Bucky
Pag
e8

  MySQL
  Serve raw/distinct queries

  Back fill

  Bucky Technology (AOL’s distributed cache &
computing framework)
  Write ahead cache: pre-compute query results and push them
into cache.
  Messaging (optional): push data directly to subscribers
  Updatesare pushed to data consumers or browsers via AIM
Complex.
  Updates go to both database and cache.

Presentation for
[CLIENT]

Architecture Diagram (over-simplified)
Pag
e9

WWW

AIM
push

Relegence

Ingestor
Distributed

Cache

Gateway
pull
WWW
Distributed

Cache
Gateway

Asset
DB

Presentation for
[CLIENT]

Data Model: SOR v.s. Query DB
Pag
e 10

  Separate query from storage to keep tables small and
query fast.
  System of Record (SOR): has all raw data
  The authoritative data store; designed for data storage
  Normalized schema: for simple key look-up; no table join.

  Query DB – de-normalized for query speed
  avoid JOIN, reduce # of trips to DB, increase throughputs.

  Read/write small chunk of data at a time so database
can get requests out quickly and process more.
  Use replication to achieve linear scalability for read.

Presentation for
[CLIENT]

Design Strategies: partitioning (Why)
Pag
e 11

  Dataset too big to fit on one host
  Performance consideration: divide and conquer
  Write: more masters (Nx) to take writes

  Read: smaller tables + more (NxM) slaves to handle read.

  Fault tolerance – distribute the risk and reduce the
impact of system failure
  Easier Maintenance – size does matter
  Faster nightly backup, disaster recovery, schema change, etc.
  Faster optimization –need optimization to reclaim disk space
after deletion, rebuild indexes to improve query speed.

Presentation for
[CLIENT]

Design Strategies: partitioning (How)
Pag
e 12

  Partition on most used keys (look at query patterns)
  Document table – on document ID

  Entity table – on entity ID

  Simple hash on IDs – no partition map; thus no
competition of read/write locks on yet another table
  Managing growth: add another partition set
  New documents are written into both old and new partition
sets for a few weeks. Then, stop writing into the old partitions.
  Queries go to the new partitions first and then the old ones if
in-sufficient results found.
  Works great in our case but might not for everyone.
Presentation for
[CLIENT]

Schema design: De-normalization
Pag
e 13

  Make query tables small:
  put only essential attributes in the de-normalized tables

  store long text attributes in separate tables.

  De-normalization: how to store and match attributes
  Single value attributes (1:1) : document ID, short string, date
time, etc. – one column, one row.
  Multi-value attributes (1:many): tricky but feasible
  Use multiple rows with composite index/key: (c1, c2, etc.)
  One row one column: CSV string, e.g., “id1, id2, id3” – SQL: “val
like ‘%id2%’”
  One row but multiple columns, e.g., group1, group2, etc. – SQL:
group1=val1 OR group2=val2 ...

Presentation for
[CLIENT]

Tips for indexing
Pag
e 14

  Simple key – for metadata retrieval
  Composite key – find matching documents
  Start with low cardinality and most used columns

  Order matter: (c1, c2, c3) != (c2, c3, c1)

  InnoDB – all secondary indexes contain primary key
  Make primary key short to keep index size small

  Queries using secondary index references primary key too.

  Integer v.s. String – comparison of numeric values is
faster => index hash values of long string instead.
  Index length – title:varchar(255) => idx_title(32)
  Enforce referential integrity on application side.
Presentation for
[CLIENT]

MySQL configuration
Pag
e 15

  Storage engine: InnoDB – row level locking
  Table space – one file per table
  Easier to maintain (schema change, optimization, etc.)

  Character set: ‘UTF-8’
  Disable persistent connection (5.0.x)

  skip-character-set-client-handshake

  Enable slow query log to identify bad queries.
  System variables for memory buffer size
  innodb_buffer_pool_size: data and indexes

  Sort_buffer_size, max_heap_table_size, tmp_table_size

  Query cache size=0; tables are updated constantly
Presentation for
[CLIENT]

Runtime statistics (per server)
Pag
e 16

  Average write rate:
  daily: < 40 tps

  max at 400 tps during recovery

  Perform best when write rate < 100 tps

  Query rate: 20~80 qps
  Query response time – shorter when indexes and
data are in memory
  75%: ~3 ms when qps < 15; ~2 ms when qps ~= 60
  95%: 6~8 ms when qps < 15; 3~4 ms when qps ~= 60

  CPU Idle %: > 99%.

Presentation for
[CLIENT]

Pag
e 17

Presentation for
[CLIENT]

Deployment Topology Consideration
Pag
e 18

•  Minimum configuration: host/DC redundency
•  DC1: host 1 (master), host 3 (slave)

•  DC2: host 2 (failover master), host 4 (slave)

•  Data locality: significant when network latency is a
concern (100 Mbps)
•  3,000 qps when DB is on remote host.
•  15,000 qps when DB is on local host.
•  Linking dependent servers across data centers
•  Push cross link up as far as possible (Topology 3): link to
dependent servers in the same data center.

Presentation for
[CLIENT]

Deployment Topology 1: minimum config
Pag
e 19
Date Center 1

DB DB

Data WWW
Consumer

DB DB

Date Center 2

Presentation for
[CLIENT]

Topology 2: link across DCs (bad)
Pag
e 20

Data
DB V V
DB Consumer
I I
P P
Data
DB Consumer G
S
L WWW
GSLB
B

Data
DB V V
Consumer
I I
DB P P
Data
DB
Consumer

Presentation for
[CLIENT]

Topology 3: link to same DC (better)
Pag
e 21

Data
DB V V
DB Consumer
I I
P P
Data
DB Consumer G
S
L WWW
B

Data
DB V V
Consumer
I I
DB P P
Data
DB
Consumer

Presentation for
[CLIENT]

Topology 4: use local UNIX socket
Pag
e 22

Data
DB V
DB Consumer
I
P
Data
DB Consumer G
S
L WWW
B

Data
DB Consumer V
I
DB P
Data
DB
Consumer

Presentation for
[CLIENT]

Production Monitoring
Pag
e 23

  Operational Monitoring: logcheck, Scout/NOC alert,
etc.
  DB monitoring on replication failure, latency, read/
write rate, performance metrics.

Presentation for
[CLIENT]

Metrics Collection
Pag
e 24

  Graphing collected metrics: visualize and collate
operational metrics.
  Help analyzing and fine tuning server performance.
  Help trace production issues and identify point of failure.

  What metrics are important?
  Host: CPU, MEM, disk I/O, network I/O, # of processes, CPU
swap/paging
  Server: Throughputs, response time

  Comparison: line up charts (throughputs, response
time, CPU, disk i/o) in the same time window.

Presentation for
[CLIENT]

Pag
e 25

Presentation for
[CLIENT]

Pag
e 26

Presentation for
[CLIENT]

Pag
e 27

Presentation for
[CLIENT]

Tuning and Optimizing Queries
Pag
e 28

  Explain: mysql> explain SELECT ... FROM …
  Watch out for tmp table usage, table scan, etc.
  SQL_NO_CACHE
  MySQL Query profiler
  mysql> set profiling=1;

  Linux OS Cache: leave enough memory on host
  USE INDEX hint to choose INDEX explicitly
  use wisely: most of the time, MySQL chooses the right index
for you. But, when table size grows, index cardinality might
change.

Presentation for
[CLIENT]

Important MySQL statistics
Pag
e 29

  SHOW GLOBAL STATUS…
  Qcache_free_blocks

  Qcache_free_memory

  Qcache_hits

  Qcache_inserts

  Qcache_lowmem_prunes

  Qcache_not_cached

  Qcache_queries_in_cache

  Select_scan

  Sort_scan

Presentation for
[CLIENT]

Important MySQL statistics (cont.)
Pag
e 30

  Table_locks_waited
  Innodb_row_lock_current_waits

  Innodb_row_lock_time

  Innodb_row_lock_time_avg

  Innodb_row_lock_time_max

  Innodb_row_lock_waits

  Select_scan

  Slave_open_temp_tables

Presentation for
[CLIENT]

Heuristic Query Optimization Algorithm
Pag
e 31

  Primary for complex cluster queries: find latest N
topics and related stories.
  Strategy: reduce the number of records database
needs to load from disk to perform a query.
  Pick a default query range. If in-sufficient docs are returned,
expand query range proportionally.
  If none return => sparse data => drop the range and retry.

  Save query range for future references.

  Result: reduce number of rows needed to process
from millions to hundreds => cut query time down
from minutes to less than 10 ms.
Presentation for
[CLIENT]

Query
range

Cluster
query

look
up

NumOfTripToDB
=0

no

Has query Use default
range? range
Compute docs to range ratio and
prorate it to a range that would return
sufficient amount of docs.

Bound query with the
range and send it to
DB yes

NumOfTrip
ToDB
>=2?

NumOfTripToDB++

Suf@icient
yes

results
numOfResults Send original
from
== 0? query to DB
query

engine?

Query

Engine

yes

Compute docs to range
ratio and save it back Return query
to the look up table for results to clients.
future use.
Presentation for
[CLIENT]

Lessons Learned
Pag
e 33

  Always load test well ahead of launch (2 weeks) to
avoid fire drill.
  Don’t rely on cache solely. Database needs to be able
to serve reasonable amount of queries on its own.
  Separate cache from applications to avoid cold start.
  Keep transaction/query simple and return fast.
  Avoid table join; limit it to 2 if really needed.
  Avoid stored procedure: results are not cached; need
DBA when altering implementation.

Presentation for
[CLIENT]

Lessons Learned (cont.)
Pag
e 34

  Avoid using ‘offset’ in LIMIT clause; use application
based pagination instead.
  Avoid ‘SQL_CALC_FOUND_ROWS’ in SELECT
  If possible, exclude text/blob columns from query
results to avoid disk I/O.
  Store text/blob in separate table to speed up backup,
optimization, and schema change.
  Separate real time v.s. archive data for better
performance and easier maintenance.
  Keep table size under control ( < 100 GB) ; optimized
periodically.
Presentation for
[CLIENT]

Lessons Learned (cont.)
Pag
e 35

  Put SQL statement (templates) in resource files so
you can tune it without binary change.
  Set up replication in dev & qa to catch replication
issues earlier
  Transactional (MySQL 5.0.x) v.s. data/mixed (5.1 or above)
  Auto-increment + (INSERT.. ON DUPLICATE UPDATE…)

  Date time column: default to NOW()

  Oversized data: increase max_allowed_packet

  Replication lag: transactions that involve index update/
deletion often take longer to complete.
  Host and data center redundancy is important –
don’t put all eggs in one basket.
Presentation for
[CLIENT]

RTN 3 Redesign
Pag
e 36

  Free Text Search with SOLR
  Real time v.s. archive shards.

  1 minute latency w/o Ramdisk.

  Asset DB partitioned – 5 rows/doc -> 25 rows/doc
  Avoid (System) Virtual Machine; instead, stack high
end hosts with processes that use different system
resources (CPU, MEM, disk space, etc)
  Better network and system resource utilization – cost effective.
  Data Locality

  More processors (< 12 ) help when under load.

Presentation for
[CLIENT]

Q&A
Pag
e 37

  Questions or comments?

Presentation for
[CLIENT]

Pag
e 38

  THANK YOU !!

Presentation for
[CLIENT]

Building and deploying large scale real time news system with my sql and distributed cache mysql_conf

More Related Content

What's hot (17)

Viewers also liked (17)

Similar to Building and deploying large scale real time news system with my sql and distributed cache mysql_conf (20)

Recently uploaded (20)

Building and deploying large scale real time news system with my sql and distributed cache mysql_conf