Billy Fung

Learnings of the Week


Billy Fung / 2016-04-18

First one

So I’m going to start writing about things I’ve learned for the week. I originally toyed with the idea of doing it daily, but then I think that might become too much work and I might end up with things that I haven’t thought through. Although it would also be more of “I did this today”, “I fixed this thing that I tried yesterday”.

I think it would be better for me to dwell on what I learned, and then hopefully every Sunday I’ll write a short succinct post about it. So here goes my first one.


One of the larger problems I was tackling this week had to do with test coverage. The way that the database driven application involved integration tests, which for all tests created a scaffolding database that does not get committed to. To do this pytest along with psycopg, within the file a Postgres testdb is created, with a transaction savepoint SAVEPOINT empty then migrations/schema/objects are applied to the database. This savepoint allows for the transaction rollback to an empty database state after a test is complete. So for every test, the test would execute whatever queries are needed to build and test the db, then afterwards we get an expected state for the next test. All this is done using a single database connection, that is created when the database test scaffolding is created.

Issues arise when trying to test a function that makes use of creating another connection to the database. Because of how the tests don’t commit to a test database, this lead to having two different database connections that queried the database in different states. This is an obvious problem because the tests would always fail. So a quick fix to this, is the patch() decorator, which allowed for a database connection to be wrapped around the function to be tested. Of course this results in a very ugly looking thing:

    with patch(database_connection):

This usage looks very hack-y and probably shouldn’t ever been done aside from in testing. But within tests, all rules are off and you just have to make sure your tests are good.

Two phase commits in PostgreSQL

Another feature I learned about Postgres is the ability to have a single transaction commit to two tables, or two databases. Two phase commits are essentially a way to run conditional commits, so if either commit fails, then rollbacks are executed to both. This is a useful tool if you need to commit to say, a customer table and an account table only if the customer commit succeeds. Having both the commits within a two phase transaction allows for this behavior to be completed in a simple manner. SQLAlchemy supports this, and although documentation is a bit lacking, it does work. One thing of note, following the SQLAlchemy documentation, Session.configure(binds={User:engine1, Account:engine2}) the binds are used to correct map the corresponding SQLAlchemy class to the proper database. So if you have within engine1 4 different tables that are required for the final commit, you will need to write all four of those classes, ie. table1:engine1, table2:engine1, table3:engine1

And lastly, the configuration for the Postgres database should have two phase commits enabled.

Finding the next business day in Python

Python libraries exist for everything, and of course there are libraries to find the next business day given an input date. But sometimes when working with libraries, you might not want to use one that isn’t supported or updated often. So to find the next business day I chose to use the standard library dateutil and [holidays]( With these two libraries, one is used to obtain the of holidays, and the other is used to create a set of reoccurring rules that can be applied to find the next day. Using dateutil, you are able to create a rule that limits the days to only weekdays, and then further, you create an exception for certain dates.

import datetime
import holidays
from dateutil import rrule

def next_business_day(invoice_date):
    # create rule for only weekday occurences, starting from invoice date
    r = rrule.rrule(
        byweekday=[rrule.MO, rrule.TU, rrule.WE, rrule.TH, rrule.FR],
        dtstart=invoice_date + datetime.timedelta(days=1)
    nz_holidays = holidays.NZ(state='AUK', years=[invoice_date.year])
    next_day = rrule.rruleset()
    # exclude all the holidays
    for holiday in nz_holidays:
        next_day.exdate(datetime.datetime(holiday.year, holiday.month,
    return next_day[0].date()