License

CC BY SA

Many of the generalizations contained in this presentation are based on a methodologically flawed, self-selecting survey of Evergreen library system administrators. Others simply reflect the author’s own biases.

Evergreen library system

Mission statement

Evergreen: highly-scalable software for libraries that helps library patrons find library materials, and helps libraries manage, catalog, and circulate those materials, no matter how large or complex the libraries.

Open-source (GPL2+): http://evergreen-ils.org

If "Libraries are the beating heart of a (community|university)", PostgreSQL is in turn at the heart of libraries that run Evergreen.

  • We go a bit beyond the canonical relational example of a library database
    • Current install creates 355 tables, 96 views, > 50 functions in 23 different schemas
    • Handles hold requests, reservations, purchases and fund management, reporting, library information, staff permissions, and more…

Who is Dan Scott?

Systems Librarian at the J.N. Desmarais Library, Laurentian University in Sudbury, Ontario (a founding member of Project Conifer)

  • Employed by IBM Canada from 1998-2006 in various positions including technical writer, support, development, and product planner
  • All for DB2 for Linux, UNIX, and Windows - with a focus on Linux and open source
  • Co-author of Apache Derby: Off to the Races media/apache_derby.jpg
  • Core Evergreen developer since 2007
  • Still feel like a PostgreSQL n00b

Evergreen library adoption (2011)

media/evergreen_libraries_2011_namerica.png
media/evergreen_libraries_2011_eurasia.png

GPLS Pines

media/pines.png
  • The birthplace of Evergreen (started 2004, 1.0 in 2006)
  • 275 libraries on a single system in the state of Georgia
  • 2.6 million patrons
  • 9.6 million items
  • 18.6 million transactions / year

BC Sitka

media/sitka.png
  • 60 libraries on a single system in British Columbia

King County Library System

media/kcls.png
  • Library system surrounding Seattle, Washington
  • 1.2 million patrons
  • 3.3 million items
  • 19 million transactions / year

Project Conifer

media/conifer.png
  • 38 libraries spanning Ontario - a mix of academic and special libraries
  • 2.5 million items

Library CONSTRAINTs

Libraries are generally resource-challenged and their systems people are asked to be responsible for many software and hardware systems, not just the library system. Thus:

  • Many Evergreen system administrators have just enough skill to get the system up and keep it running
  • Despite the critical role it plays in system performance, PostgreSQL is often learned on a need-to-know basis in production
    • "All-in-one" underprovisioned server
    • Logs and data on same partition
    • Limited tuning; pg_tune or bust
    • Default statistics target at 50
    • Backups via pg_dump or incremental file system backups

It’s not all bad

  • Many sites rely on a third party company for setup and support, although too much dependency is always a concern
  • Several Evergreen system administrators at PGcon this year; collectively, we will be stronger (and perhaps develop a set of Evergreen-specific best practices)
  • Our development practices are maturing:
    • Code reviews are mandatory before committing to master
    • We have (some) standard sample data, unit tests, and a CI server
    • We have more documentation and broader communication
  • Opportunities for consulting and training for PostgreSQL experts; help us make Evergreen a success throughout the world, and earn a living do it :)

Horrible, horrible library data

Central element of most library data is the MARC record, a combination of fixed-length fields and variable-length fields that encodes the bibliographic description of an object.

LDR 00969cam a22002774a 4500
001 14338589
005 20070508144242.0
008 060412s2005 cc 001 0 eng c
010     ‡a 2006273753
020     ‡a9780596007591 (pbk.)
020     ‡a0596007590 (pbk.)
050 0 0 ‡aQA76.76.D47 ‡bF634 2005
082 0 0 ‡a005.1 ‡222
100 1   ‡aFogel, Karl.
245 1 0 ‡aProducing open source software :
        ‡bhow to run a successful free software project / ‡cKarl Fogel.
250     ‡a1st ed.
260     ‡aBeijing ; ‡aSebastopol, CA : ‡bO'Reilly, ‡c2005.
300     ‡axx, 279 p. ; ‡c24 cm.
500     ‡aIncludes index.
650   0 ‡aComputer software ‡xDevelopment.
650   0 ‡aOpen source software.

Mike Rylander, Evergreen’s eeevil database genius

Mike Rylander
Figure 1. Mike Rylander was sent from the future to defend the open source library system world from the tyranny of MARC

Indexing library data the Evergreen way

Generally, start with MARC (serialized as MARCXML) in biblio.record_entry.marc:

<record xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
<leader>00969cam a22002774a 4500</leader>
<controlfield tag="001">14338589</controlfield>
<controlfield tag="005">20070508144242.0</controlfield>
<controlfield tag="008">060412s2005    cc            001 0 eng c</controlfield>
<datafield tag="010" ind1=" " ind2=" ">
  <subfield code="a">  2006273753</subfield>
</datafield>
<datafield tag="020" ind1=" " ind2=" ">
  <subfield code="a">9780596007591 (pbk.)</subfield>
</datafield>
<datafield tag="082" ind1="0" ind2="0">
  <subfield code="a">005.1</subfield>
  <subfield code="2">22</subfield>
</datafield>
<datafield tag="100" ind1="1" ind2=" ">
  <subfield code="a">Fogel, Karl.</subfield>
</datafield>
<datafield tag="245" ind1="1" ind2="0">
  <subfield code="a">Producing open source software :</subfield>
  <subfield code="b">how to run a successful free software project /</subfield>
  <subfield code="c">Karl Fogel.</subfield>
</datafield>
</record>
  • Yes, XML does make it better!

Random access by field-subfield

To support a MARC expert search, we populate metabib.full_rec:

  1. source FK pointing to biblio.record_entry.id
  2. value containing normalized text
  3. index_vector index column with associated trigger
SELECT * FROM metabib.full_rec WHERE record = 884755 AND tag = '245';
-[ RECORD 1 ]+-------------------------------------------------------
id           | 22640054
record       | 884755
tag          | 245
ind1         | 1
ind2         | 0
subfield     | a
value        | producing open source software
index_vector | 'open':2 'produc':1 'softwar':4 'sourc':3

83M metabib.full_rec rows in Conifer’s production database

Challenge: some fields such as general notes are lengthy, blowing past the btree maximum.

Eventual solution: Create a SUBSTR(value, 1, 1024) expression index on metabib.full_rec, rename the table to metabib.real_full_rec, and create a view called metabib.full_rec on top of it.

Indexing title / author / subject / keyword

  1. Transform MARCXML into more human-friendly, semantic XML (generally MODS)
  2. Define index classes with weighted fields (class, field, XML transform, XPath, weight)
  3. Extract corresponding chunks into metabib.*_field_entry.value
  4. index_vector index column with associated trigger
-[ RECORD 1 ]+-------------------------------
id           | 4234610
source       | 884755
field        | 6
value        | Producing open source software
             | how to run a successful free
             | software project
index_vector | 'a':8 'free':10 'how':5 'open':2
             | 'produc':1 'project':12 'run':7
             | 'softwar':4,11 'sourc':3
             | 'success':9 'to':6

29M metabib.*_field_entry rows in Conifer’s production database

Adventures in text search: Evergreen 1.0

Circa 2006, PostgreSQL 8.0/8.1

  • Text search built on TSearch2 contrib module ca. PostgreSQL 8.0
    • Thank you Oleg and Teodor!
  • All indexed values created externally via Perl scripts, then initially loaded via COPY
    • Good for parallelized bulk loading
    • Brittle due to potential for ID conflict
    • Terrible for consistency, as updates to indexed values were managed by the application (and thus often did not happen)

Adventures in text search: Evergreen 1.6

Circa 2009, PostgreSQL 8.3/8.4

  • Integrated full text search in PostgreSQL!
    • Thank you Oleg and Teodor!
  • Still using TSearch2 contrib for compatibility
  • Revelations about LCCOLLATE and LCCTYPE:
    • Debian / Ubuntu created UTF8 clusters by default
    • Negative performance impact on search was obfuscated until a real set of data is loaded

Adventures in text search: Evergreen 2.0

Circa 2011, PostgreSQL 9.0

  • Evolved to database functions (plperlu, plpgsql, SQL) & triggers for indexing and updates, avoiding ID conflicts and improving consistency
    • Trigger applies a series of customizable normalizations, implemented as database functions, for each value for a given field before insertion into the tsvector column
    • Search against a given field applies the same normalizations to the incoming search term(s)
  • New features for users:
    • Wildcard searches
    • Exposed the Boolean OR operator (joining NOT and AND)
      • Librarians rejoiced! Nobody else noticed :)
  • Some sites adopting GIN indexes

Adventures in text search: Evergreen 2.2

Circa 2012, PostgreSQL 9.1

  • Still installing TSearch2 contrib module (force of habit; not really required)

Bad news for text search

  • Serialized serial operations seem to be a bottleneck for bulk loading and reingesting
  • ORDER BY rank with ARRAY_AGG(DISTINCT source) kills performance for large results: 600MB merge sort for 500K hits
    • Granular index design compounds problems for general searches, requiring DISTINCT & therefore disk-based sort due to outlandish memory demands
    • Good news: many nights of EXPLAIN ANALYZE later, committed a change yesterday that improves performance significantly (in at least one environment): CTE and avoidance of ARRAY_AGG(DISTINCT source)
  • Stemming - desired, used, but problematic for academics and their multilingual collections in our implementation
  • Stop words are not an option:
    • or is gold to a university that focuses on mining
    • It is a popular novel
    • The The is a band

Outsource to Solr?

Solr comes up as an option for sub-second results:

  • Broader adoption throughout library development community
  • Perceived as having more mature and diverse analyzers / tokenizers / token filters
  • Several branches exist for synchronizing Evergreen contents with a Solr instance

However, convenience and consistency of having full-text search managed by PostgreSQL generally outweighs perceived advantages of Solr.

Still not fun explaining this advantage to users and staff when their overly general query simply times out.

Functions / stored procedures

  • Integral to indexing and search
    • Custom functions sometimes required to overcome PostgreSQL limitations
    • LOWER() on Unicode strings insufficient; thus we use plperlu to invoke lc()
  • Similarly, increasingly embedding heavy lifting into the database
  • Borrowing periods, fines, and other policies based on the complex matrix of borrower, item, and library attributes that libraries demand
  • All custom routines written in SQL, plpgsql, or plperlu
    • Recently started tweaking default attributes like COST, ROWS, and IMMUTABLE/STABLE/VOLATILE for performance purposes
    • GSoC student will be hunting bottlenecks that can be addressed via rewrites in SQL or C
    • Adoption of new native functions like STRING_AGG() vs. ARRAY_TO_STRING(ARRAY_AGG()) and rewriting connectby() as WITH RECURSIVE CTEs

Active tables

The bibliographic record table is one of the more active tables in our schema:

biblio.record_entry triggers
a_marcxml_is_well_formed BEFORE INSERT OR UPDATE
a_opac_vis_mat_view_tgr AFTER INSERT OR UPDATE
aaa_indexing_ingest_or_delete AFTER INSERT OR UPDATE
audit_biblio_record_entry_update_trigger AFTER DELETE OR UPDATE
b_maintain_901 BEFORE INSERT OR UPDATE
bbb_simple_rec_trigger AFTER INSERT OR DELETE OR UPDATE
c_maintain_control_numbers BEFORE INSERT OR UPDATE
fingerprint_tgr BEFORE INSERT OR UPDATE

Debian/Ubuntu packaging

  • Most Evergreen sites rely on packages and don’t have expertise
    • Therefore Martin Pitt’s backports are a godsend
  • But packaging decisions introduce well-known compatibility pain points as well
    • Conflicting approaches to starting / stopping clusters
    • Location of configuration files
    • Upgrade challenges (pg_upgrade vs pg_upgradecluster)

Materialized views

For reporting simplicity and increased performance, materialized views (AKA materialized query tables) rock

  • We fake materialized views using triggers and rules—but occasionally get things subtly wrong
    • A mistake with money.materialized_billable_xact_summary was painful, because it lead to patrons expecting refunds they weren’t owed
  • Would love to have CREATE TABLE or CREATE VIEW options for REFRESH IMMEDIATE and REFRESH DEFERRED that would do the work for us
  • Also, would love a pony

Hstore

Currently using hstore effectively in two places:

  • Single-valued fields
    • Bibliographic record attributes that can have only one instance per record (such as year of publication)
    • Even though there are already many of them, librarians seem to continually spawn new record attributes
  • Function arguments: avoids torturous variations of the same function definition with different signatures
    • For example, specifying different levels of limits: unapi.bre(..., 'acn=>5,acp=>10')
  • It works!

Connection pooling

Would like to implement connection pooling to reserve server resources for core database processes

  • (Local anecdote): pgpool-II failed in production after a few hours with a hard lockup
    • Could be a packaging issue; didn’t have time to dig further
    • Only one site is still running pgpool successfully
  • Plan to investigate pgBouncer

Replication

  • Slony has been the go-to option for reporting replicas
    • Limitations on commands such as TRUNCATE have bitten us, as developers typically don’t test in a Slony environment
  • WAL archiving / log shipping has been the go-to option for backup and disaster recovery, but many moving parts and options were daunting
  • Streaming replication is simple to set up and great for disaster recovery
    • However, in a naive implementation (ours at Conifer), many queries time out
    • Will be looking into vacuum_defer_cleanup_age and hot_standby_feedback thanks to Phillip Sorber’s replication tutorial

Inheritance

  • Used sparingly but effectively for modelling objects with similar behaviour
    • Things like copies of books (asset.copy is a parent of serial.unit)
    • Transactions that might have costs attached (action.circulation is a child of money.billing)
  • Occasionally stab ourselves by forgetting triggers, unique / FK / PK constraints (or having to customize them to be more flexible)

Schema evolution

  • Evergreen has no automated solution for creating point-to-point upgrades
    • Currently, we write serially incrementing upgrade scripts that get concatenated & munged at release time
  • (9.2) Avoiding table rewrites when we add a column with a default value will be appreciated
  • DISABLE TRIGGER ALL helps performance, when we remember and when appropriate

Upgrading PostgreSQL

Libraries are generally averse to frequent system change, for the usual business reasons (avoiding downtime, risk and retraining).

  • One Evergreen upgrade per year is about right
  • Generally prefer to avoid upgrading distributions or major components (such as PostgreSQL) at the same time
    • Thank you for your generous support policies; many libraries will be jumping from 8.4 to 9.1 in the next six months
  • Definitely want to avoid downtime; with rise in electronic resources, libraries are 24x7 businesses
    • pg_dump / pg_restore cycle was a bit painful, even with parallel restore
    • pg_upgrade definitely helps; 148 minutes for a 90 GB database
      • Not yet integrated into Debian/Ubuntu packagers' pg_upgradecluster, which does a full dump/restore

Kudos to PostgreSQL

  • PostgreSQL has never been responsible for Evergreen data loss
  • PostgreSQL has never been a bottleneck for Evergreen operations, for the largest and busiest of Evergreen sites (outlier queries excepted)
  • Thorough documentation: release notes, core docs, active community of bloggers
  • Supportive, welcoming community (#postgresql, mailing lists))
  • Continual improvement and evolution

Help us with our mission

Mission statement

Evergreen[:] highly-scalable software for libraries that helps library patrons find library materials, and helps libraries manage, catalog, and circulate those materials, no matter how large or complex the libraries.

  • PostgreSQL is close to your heart, and it’s at the heart of Evergreen
    • Help bring Evergreen (and PostgreSQL) to a library near you
    • Make our heart beat faster!