This repo implements distributed transaction manager using Snapshot Sharing mechanism. General concepts and alternative approaches described in postgres wikihttps://wiki.postgresql.org/wiki/DTM.

Backend-DTM protocol description can be found in[dtmd/README](dtmd/README).

* Patch postgres using xtm.patch. After that build and install postgres in usual way.

cd~/code/postgres

patch -p1<~/code/pg_dtm/xtm.patch

* Install pg_dtm extension.

export PATH=/path/to/pgsql/bin/:$PATH

cd~/code/pg_dtm

make&& make install

* Run dtmd.

cd~/code/pg_dtm/dtmd

make

mkdir /tmp/clog

./bin/dtmd&

* To run something meaningful you need at leat two postgres instances. Also pg_dtm requires presense in```shared_preload_libraries```.

initdb -D ./install/data1

initdb -D ./install/data2

echo"port = 5433">> ./install/data2/postgresql.conf

echo"shared_preload_libraries = 'pg_dtm'">> ./install/data1/postgresql.conf

echo"shared_preload_libraries = 'pg_dtm'">> ./install/data2/postgresql.conf

pg_ctl -D ./install/data1 -l ./install/data1/log start

pg_ctl -D ./install/data2 -l ./install/data2/log start

For a cluster-wide deploy we use ansible, more details in tests/deploy_layouts. (Ansible instructions will be later)

Now cluster is running and you can use global tx between two nodes. Let's connect to postgres instances at different ports:

create extension pg_dtm;-- node1

createtableaccounts(user_idint, amountint);-- node1

insert into accounts (select2*generate_series(1,100)-1,0);-- node1, odd user_id's

create extension pg_dtm;-- node2

createtableaccounts(user_idint, amountint);-- node2

insert into accounts (select2*generate_series(1,100),0);-- node2, even user_id's

select dtm_begin_transaction();-- node1, returns global xid, e.g. 42

select dtm_join_transaction(42);-- node2, join global tx

begin;-- node1

begin;-- node2

update accountsset amount=amount-100where user_id=1;-- node1, transfer money from user#1

update accountsset amount=amount+100where user_id=2;-- node2, to user#2

commit;-- node1, blocks until second commit happend

commit;-- node2

To ensure consistency we use simple bank test: perform a lot of simultaneous transfers between accounts on different servers, while constantly checking total amount of money on all accounts. This test can be found in tests/perf.

> go run ./tests/perf/*

-C value

Connection string (repeatfor multiple connections)

-a int

The number of bank accounts (default 100000)

-b string

Backend to use. Possible optinos: transfers, fdw, pgshard, readers. (default"transfers")

-gUse DTM to keep global consistency

-iInit database

-lUse'repeatable read' isolation level instead of'read committed'

-n int

The number updates each writer (readerincase of Reades backend) performs (default 10000)

-pUse parallel execs

-r int

The number of readers (default 1)

-s int

StartID. Script will update rows starting from this value

-vShow progress and other stufffor mortals

-w int

The number of writers (default 8)

go run ./tests/perf/*.go \

-C"dbname=postgres port=5432" \

-C"dbname=postgres port=5433" \

-g -i

go run ./tests/perf/*.go \

-C"dbname=postgres port=5432" \

-C"dbname=postgres port=5433" \

-g

We also provide a patch, that enables support of global transactions with postres_fdw. After patching and installing postres_fdw it is possible to run sametest via fdw usig key```-b fdw```.

pg_shard.use_dtm_transactions = 1

The testing process involves multiple modules that perform different tasks. The

modules and their APIs are listed below.

Governs the whole testing process. Runs different workloads during different

troubles.

Puts workloads against the database. Writes logs that are later used by

`valideaux`.

*`start(id, workload, cluster)` - starts a`workload` against the`cluster`

and call it`id`.

*`stop(id)` - stops a previously started workload called`id`.

Manages the database nodes.

*`deploy(driver, ...)` - deploys a cluster using the specified`driver` and

other parameters specific to that driver. Returns a`cluster` instance that is

used in other methods.

* `

This is the troublemaker that messes with the network, nodes and time.

*`cause(cluster, trouble)` - causes the specified`trouble` in the specified

`cluster`.

*`fix(cluster)` - fixes all troubles caused in the`cluster`.

Validates the logs of stresseaux.

Generates reports on the test results. This is usually a table that with

`trouble` vs`workload` axes.