Migration From DB2 in a Large Public Setting: Lessons Learned

Balázs Bárány and Michael Banck

PGConf.EU 2017
Introduction

I Federate state ministry in Germany

I Hosting by state’s central IT service centre
I Michael worked as an external consultant for both
I Balázs took over DBA role and migration lead at ministry
I Michael continues to support the service centre’s Postgres operations

Balázs Bárány and Michael Banck PGConf.EU 2017 1 / 20

Introduction

I Proof-of-Concept version of this talk presented at pgconf.eu 2015

I Slides still available on https://wiki.postgresql.org/
I DB2 UDB is the z/OS mainframe edition of IBM’s DB2 database
I DB2 UDB central database and application server (“the Host”) in German state ministry
I Used by programs written in (mostly) Software AG Natural and Java (some PL/I)
I Natural (and PL/I) programs directly executed on the mainframe, no network round-trip
I Business-critical, handles considerable payouts of EU subsidies
I Crunch-Time in spring when users apply for subsidies

Balázs Bárány and Michael Banck PGConf.EU 2017 2 / 20

Prior Postgres Usage

I Postgres introduced about 12 years ago due to geospatial requirements (PostGIS,

nothing comparable for DB2 at the time)
I Started using Postgres for smaller, non-critical projects about 7 years ago
I Modernized the software stack merging geospatial and business data about 5 years ago
I In-house code development of Java web applications (Tomcat/Hibernate)
I Business-logic in the applications, almost no (DB-level) foreign keys, no stored
procedures
I Some business data retrieved from DB2, either via a second JDBC connection, or via
batch migrations
I Migrated all Natural and PL/I programs to Java/Postgres

Balázs Bárány and Michael Banck PGConf.EU 2017 3 / 20

Application Migration Strategy

I Java Applications
I Development environment switched to Postgres and errors fixed
I Not a lot of problems if Hibernate is used
I Potentially migrated to modernized framework
I PL/I Applications
I Rewritten in Java (only a few)
I Natural Applications
I Automatic migration/transcription into (un-Java, but correct) Java on DB2
I Test of migrated “Java” application on the original data
I Test on schema migrated to PostgreSQL
I Multi-year project facilitated by an external consultancy

Balázs Bárány and Michael Banck PGConf.EU 2017 4 / 20

Setup Before the Migration

I Postgres
I Postgres-9.4/PostGIS-2.1 (upgrade to 9.6/2.3 planned in late 2017)
I SLES11, 64 cores, 512 GB memory, SAN storage
I HA 2-node setup using Pacemaker, two streaming standbys (one disaster recovery
standby)
I Roughly 1.3 TB data, 22 schemas, 440 tables, 180 views in PROD
I Almost no stored procedures (around 10)
I DB2
I DB2 UDB Version 10
I Roughly 600 GB data in PROD instance
I Almost no stored procedures (around 20, written in PL/I)

Balázs Bárány and Michael Banck PGConf.EU 2017 5 / 20

Steps towards migration

I Natural migration to Java delayed

I Originally planned for November 2015, ready in July 2017
I Gave us one year for testing the migration process
I Several Java projects maintained by external developers have been (mostly)
successfully tested on local Postgres deployments
I First production migration of a complex Java program and its schema done in early
2016
I Required daily migration of core tables (DB2 to PostgreSQL) starting at that point
I Separate DB2 database operated by the ministry migrated from mainframe to another
Postgres instance successfully in Q1/2017
I Just a data migration, schema was migrated by hand

Balázs Bárány and Michael Banck PGConf.EU 2017 6 / 20

Tools used for the migration

I SQLWorkbench/J (http://www.sql-workbench.net) v117.6

I Java-based, DB-agnostic workbench GUI
I Heavily-used in-house already, installed on workstations
I Allows for headless script/batch operation via internal programs
I Used for schema migration and data export from DB2
I pgloader (http://pgloader.io) 3.2.0
I Postgres bulk loading and migration tool written in Lisp
I Open Source (PostgreSQL license)
I Written and maintained by Dimitri Fontaine (PostgreSQL major contributor)
I Used for data import into Postgres

Balázs Bárány and Michael Banck PGConf.EU 2017 7 / 20

Schema-Migration

I General Approach
I Dump schema objects into an XML representation
I Transform XML into Postgres DDL via XSLT
I Provide compatibility environment for functions called in views and triggers
I Post-process SQL DDL to remove/work-around remaining issues
I Handle triggers separately
I Ignore functions/stored procedures (out-of-scope)

Balázs Bárány and Michael Banck PGConf.EU 2017 8 / 20

DB2 Compatibility Layer (db2fce)

I Similar (in spirit) to orafce, only SQL-functions so far

I https://github.com/credativ/db2fce, PostgreSQL license
I SYSIBM.SYSDUMMY1 view (similar to Oracle’s DUAL table)
I SELECT 1 FROM SYSIBM.SYSDUMMY1;
I db2 Schema:
I Time/Date: MICROSECOND()/SECOND()/MINUTE()/HOUR()/DAY()/MONTH()/
YEAR()/DAYS()/MONTHS BETWEEN()
I String: LOCATE()/TRANSLATE()/STRIP()
I Casts: CHAR()/INTEGER()/INT()/DOUBLE()/DECIMAL()/DEC()
I Aliases: VALUE() (for coalesce()), DOUBLE (for DOUBLE PRECISION type), ^= (for <>
/ != operators), !! (for || operator)
I search path changed to ’db2, public’ in database configuration
Balázs Bárány and Michael Banck PGConf.EU 2017 9 / 20
Data Migration, Encountered Problems

I Several tables had \x00 values in them, resulting in "invalid byte sequence for
encoding UTF8: 0x00" errors
I Exporting tables with a column USER resulted in WbExport writing the username of
the person running it
I Default timestamp resolution was too coarse, leading to duplicate key violations
I NUMERIC(X,Y) columns were exported with a precision of 2 only
I Import of timestamps invalid in daylight saving change time rejected by PostgreSQL
I Export them with -Duser.timezone=GMT despite being local (Central European)
timestamps
I Objects in target DB with the same name as in the source, but different contents
I Renamed in source system

Balázs Bárány and Michael Banck PGConf.EU 2017 10 / 20

Full Migration

I Closed databases for “normal” usage

I Source DB switched to read only
I PostgreSQL: removed USAGE on schemas from non-DBA users
I Notified users with open connections
I Deactivated HA watchdog, disaster recovery
I Scripted (automatic) migration process:
I Dumped schema to XML, converted to DDL, post-processed
I Dropped indexes, constraints and triggers
I Exported data
I Imported data
I Set sequence values
I Created indexes, constraints and triggers
I Created grants

Balázs Bárány and Michael Banck PGConf.EU 2017 11 / 20

Full Migration, Results

I Full migration in 3 processes (different schemas) in 14 hours incl. index building

I Database was gaining 1 GB every 2 minutes when 3 processes were writing
I Up to 80 Mbit/sec both incoming and outgoing on the network interfaces
I Up to 120 Mbit/sec when writing (4 CPUs at the limit)
I Data validation jobs started whenever a schema was ready
I Minimal differences in floating point representation
I Everything else identical, including binary data and sequence values
I Watching logs for errors while the applications start
I A few schema or table permissions were missing
I Tables missing
I Dropped from source system before, application was not tested

Balázs Bárány and Michael Banck PGConf.EU 2017 12 / 20

SQL Differences

I Migration Guide in PostgreSQL wiki

I https://wiki.postgresql.org/wiki/File:DB2UDB-to-PG.pdf
I Age and Author unknown
I Noticed SQL Differences
I CURRENT TIMESTAMP etc. (but CURRENT TIMESTAMP is supported by DB2 as well)
I Casts via scalar functions like INT(foo.id)
I CURRENT DATE + 21 DAYS
I ‘2100-12-31 24.00.00.000000’ timestamp in data - year 2100/2101
I Operators like != instead of <>
I “Default default” values: attribute INTEGER DEFAULT
I Like attribute INTEGER DEFAULT 0 in PostgreSQL

Balázs Bárány and Michael Banck PGConf.EU 2017 13 / 20

Behaviour Differences

I DB2 sorts data by GROUP BY keys, no need for ORDER BY

I PostgreSQL doesn’t guarantee this
I Sorting differences
I EBCDIC: numbers after characters (ASCII: before)
I EBCDIC: special characters after characters and numbers
I Similar behaviour with C collation in Postgres
I Applications using ECBDIC order inside values
I Application got “duplicate key value” error
I Tried to use CURRENT TIMESTAMP as primary key
I Postgres: Start of transaction. DB2: current time regardless of transaction.
I Application trying to insert NULL into field with DEFAULT
I DB2 accepted the NULL and used the DEFAULT
I For Postgres, we had to create a trigger to fix the INSERTs
Balázs Bárány and Michael Banck PGConf.EU 2017 14 / 20
Transaction handling

I Some queries in DB2 using WITH UR

I UR = Uncommitted Read
I Performance optimization to avoid locks (but getting inconsistent data)
I No comparable built-in mechanism in PostgreSQL, but locking not a huge problem
I PostgreSQL cancels the entire transaction after an error, ROLLBACK necessary
I Some program logic needed changes (bad error handling, errors used for branching logic)

Balázs Bárány and Michael Banck PGConf.EU 2017 15 / 20

Performance

I Most applications comparable or faster after migration

I A few with mainframe access patterns slower
I Some smaller impacts: different indexing, JDBC oddities, . . .
I A few huge problems (application not usable)
I Indexes correct, query is fast in SQL client
I Not using “best” index when called from prepared query
I Found the reason: ID (integer) field was queried with NUMERIC parameter

Balázs Bárány and Michael Banck PGConf.EU 2017 16 / 20

JDBC: defaultRowFetchSize

I Applications started crashing with Out of Memory errors

I Pattern: SELECT * FROM <big table>, read some rows for display
I PostgreSQL JDBC reads the whole data set by default
I DB2 didn’t, so the application was working fine
I Solution: setting defaultRowFetchSize to a reasonable value (e. g. 10000)
I No negative effects (negligible performance hit?)

Balázs Bárány and Michael Banck PGConf.EU 2017 17 / 20

JDBC: stringtype

I Errors with prepared statements, but query works in SQL client

I This works (automatic casting to date):
SELECT * FROM t WHERE dat = ’2017-08-01’;
I This doesn’t (with param1 = ’2017-08-01’):
SELECT * FROM t WHERE dat = ?;
I Could affect date, timestamp, numeric and Boolean columns
I Comparison of date type with “forced” text type fails, no automatic cast
I Solution: stringtype=unknown
I Fine for the affected applications
I Might be wrong in some situations, e. g. garbled date format

Balázs Bárány and Michael Banck PGConf.EU 2017 18 / 20

Summary

I Customer is happy
I Big savings on mainframe costs
I mainframe performance units, DB2, Natural runtime, . . .
I One modern database for business and GIS data, pleasant usage
I Better standards for DB roles and permissions, change management etc.

Balázs Bárány and Michael Banck PGConf.EU 2017 19 / 20

Contact

I DB2 compatibility extension: https://github.com/credativ/db2fce

I Michael Banck <[email protected]>
I http://www.credativ.de/postgresql-competence-center
I Balázs Bárány <[email protected]>
I https://datascientist.at/

Questions?

Balázs Bárány and Michael Banck PGConf.EU 2017 20 / 20