----
presentation_topic: GitDB
presentation_title: GitDB - plumbing for databases
presentation_place: GitTogether @ GooglePlex, Mountain View
presentation_date: October 29th, 2008

----
== Git as a database back-end

* motivation
* design
* discussion

----
== GitDB design elements

* row format
* information_schema (DDL) schema
* table to tree layout
* ACID protocol
* git plumbing commands to do relational things

----
Row format

* completely arbitrary, really.  could even be mixed.
* suggested formats:
** Pg binary row format (minus unnecessary headers)
** TSV + text encoded values

----
information_schema

* required for DDL transactions
* allows types to be omitted from row format
* suggested formats:
** Pg pg_class tables
** SQL standard information_schema
** MySQL - not as complete/flexible as Pg
* should support basic contstraints (FK, unique, expression)

----
table to tree layout

* top level as relation name
* next level is the text form of primary key
* Index-Organised-Table is therefore the default

----
table to tree layout 2

* object as a session of writing row data, for fast bulk insert/full scan
* implies object names become range of rows within
* generally tables will use (specially selected) UUID format so that they can write many rows without worrying about key overlap
* some tree entries will denote a deletion or replacement of a row

----
table to tree layout 3

* may be split into sub-trees for b-tree indexing large tables
* Indexes are discrete entities and contain the primary key of the referent row

----
ACID protocol

* log the key range of locks acquired for the transaction in the commit message
* writers can then check that objects have not changed, using index ID matching, or (worst case) checking rows didn't change
* highest SQL isolation level supported already
* could support full predicate matching for specialist applications by acquiring locks on temporary views

----
Git commands

* basic operations:
** scanner/random seeker - git-db-select
** inserter/updater - git-db-upsert
** deleter - git-db-delete
* tx logic:
** git-db-txn ( begin | commit )
** active transaction makes above commands spool activity
** git-db-txn merge commit
* DDL operations:
** create relation - git-db-ddl-create
** alter relation - git-db-ddl-alter
* complex query machinery (later):
** join select streams - git-db-join
** evaluate expressions on select streams - git-db-expr
** subselect - git-db-loop
** SQL parser/query planner - git-db-sql

----
Potential uses

* solid foundation for distributed applications
* auditing/multi-master applications
* might even be faster/smaller DBs than existing databases
* give Progres experts a new feeling of home