® npr.org Migration From Oracle to MySQL An NPR Case Study By Joanne Garlow.

®

npr.org

Migration From Oracle to MySQL

An NPR Case Study

By Joanne Garlow

Overview

Background

Database Architecture

SQL Differences

Concurrency Issues

Useful MySQL Tools

Encoding Gotchas

Background NPR (National Public Radio)

Leading producer and distributor of radio programming All Things Considered, Morning Edition, Fresh Air, Wait,

Wait, Don’t Tell Me, etc. Broadcasted on over 800 local radio stations nationwide

NPR Digital Media Website (NPR.org) with audio content from radio programs Web-Only content including blogs, slideshows, editorial

columns About 250 produced podcasts, with over 600 in directory Mobile apps and sites Syndication

High-Level System Architecture

Limitations of the Oracle Architecture

Reached capacity of single system to support

our load

Replication outside our budget

Databases crashes were becoming frequent

Database Architecture Goals

Redundancy

Scalability

Load balancing

Separation of concerns

Better security

High-Level System Architecture

• Updated by our Content Management System• Transaction Oriented • Resource Contention• Highly Normalized

• Isolation from main website • Read-only by our webservers• Horizontally scalable

Database Architecture

MainInnoDB

AMGMyISAM

PUBLICInnoDB

STATIONSInnoDB

MainRO slave

Main RO slave

Content Mgmt System

Web Servers

Scripts BackupRO slave

• Read and updated only by our website• Low resource contention• Small tables or log tables• Short Transactions

• Updated by a nightly script• Read-only by our Content Management System• Need fast full text queries (replacing Oracle Text)• Large tables

• Updated by a quarterly script• Read-only from our website• Some log type information written• Low resource contention• No transactions

Issues When Converting SQL MySQL is case sensitive

Oracle outer join syntax (+) -> OUTER JOIN clause

Oracle returns a zero to indicate zero rows updated – MySQL

returns TRUE (1) to indicate it successfully updated 0 rows

MySQL sorts null to the top, Oracle sorts null to the bottom

Use “order by – colName desc” for sorting asc with nulls at

bottom

MySQL has Limit clause – YAY!

No sequences - DOH! Continued….

Replacing Oracle Sequences Initialize a table with a single row:

CREATE TABLE our_seq (

id INT NOT NULL

);

INSERT INTO our_seq (id) VALUES (120000000);

Do the following to get the next number in the “sequence”:

UPDATE our_seq SET id=LAST_INSERT_ID(id+1);

SELECT LAST_INSERT_ID();

Replacing Oracle Sequences For updating many rows at once, get the total number of unique IDs

you need first:SELECT @totalRows := COUNT(*) FROM...

Then update npr_seq by that many rows:UPDATE npr_seq SET id=LAST_INSERT_ID(id+@totalRows);

and store that ID into another variable:SELECT @lastSeqId := LAST_INSERT_ID();

Then use the whole rownum workaround described above to get a unique value for each row:INSERT INTO my_table (my_primary_id . . . ) SELECT @lastSeqId - (@rownum:=@rownum+1), . . . FROM (SELECT @rownum:=-1) r, . . .

Converting Functions NVL() -> IFNULL() or COALESCE() DECODE() -> CASE() or IF() Concatenating strings || -> CONCAT()

‘test’ || null returns ‘test’ in Oracle CONCAT(‘test’,null) returns null in MySQL

LTRIM and RTRIM -> TRIM() INSTR() works differently.

Use LOCATE() for Oracle’s INSTR() with occurrences = 1.

SUBSTRING_INDEX() and REVERSE() might also work.

Converting Dates

sysdate -> now()

Adding or subtracting In Oracle “– 1” subtracts a day

In MySQL “- 1” subtracts a milisecond – must use

“interval”

TRUNC() -> DATE()

TO_DATE and TO_CHAR -> STR_TO_DATE and

DATE_FORMAT

Update Differences You can't update a table that is used in the

WHERE clause for the update (usually in an "EXISTS" or a subselect) in mysql.  UPDATE tableA SET tableA.col1 = NULL WHERE tableA.col2 IN (SELECT tableA.col2 FROM tableA A2, tableB WHERE tableB.col3 = A2.col3 AND tableB.col4 = 123456);

You can join tables in an update like this (Much easier!):UPDATE tableA INNER JOIN tableB ON tableB.col3 = tableA.col3 SET tableA.col1 = NULL WHERE tableB.col4 = 123456;

RANK() and DENSE_RANK()

We really found no good MySQL equivalent for

these functions

We used GROUP_CONCAT() with an ORDER BY

and GROUP BY to get a list in a single column

over a window of data

Collation

You can set collation at the server, database,

table or column level.

Changing the collation at a higher level (say on

the database) won’t change the collation for

preexisting tables or column.

Backups will use the original collation unless

you specify all the way down to column level.

Concurrency Issues

In our first round of concurrency testing, our

system ground to a halt! Deadlocks

Slow Queries

MySQL configuration sync_binlog = 1 // sync to disk, slow but safe

innodb_flush_log_at_trx_commit = 1 // write each

commit

transaction_isolation = READ-COMMITTED

Useful MySQL Tools

MySQL Enterprise Monitor

http://www.mysql.com/products/enterprise/

MySQL GUI Tools Bundle:

http://dev.mysql.com/downloads/gui-tools/5.0.html

MySQL Query Browser similar to Oracle’s SQL

Developer

MySQL Administrator

Innotop and innoDB Status

innotop

http://code.google.com/p/innotop

Helped us identify deadlocks and slow queries

(don’t forget the slow query log!)

In mysql, use

show engine innodb status\G;

Useful for contention and locking issues

Query Profiling

Try the Query Profiler with Explain Plan when

debugging slow queries

http://dev.mysql.com/tech-resources/articles/using-

new-query-profiler.html

Concurrency Solution

Tuning our SQL and our server configuration

helped

Turns out that the RAID card we were using

had no write cache at all. Fixing that allowed

us to go live.

Encoding Gotcha’s

Switched from ISO-8859-1 to UTF-8

Migration Tool Issues with characters that actually were not

ISO-8859-1 in our Oracle database

Lack of documentation for the LUA script

produced by the migration GUI

Update encoding end to end JSPs, scripts (Perl), PHP, tomcat (Java)

Continuing Issues

Bugs with innodb locking specific records (as

opposed to gaps before records) Uncommitted but timed out transactions

Use innotop or “show engine innodb status\G; “

and look for threads waiting for a lock but no

locks blocking them

Requires MySQL reboot

Questions? Joanne Garlow

jgarlow@npr.org http://www.npr.org/blogs/inside