Transcript

The video above was recorded as the Day 2 keynote at Ruby Kaigi on September 9th, 2016.

Background

I’ve been occasionally criticized since my Make Ruby Great Again talk that it’s been a while since I’d contributed much to the Ruby community. I was thinking about this when I was asked to design a talk for Ruby Kaigi (Japan’s national Ruby conference) that was about something other than testing or feelings. (This forced me to consider whether I knew anything other than testing and feelings.)

I sat on the idea for a while before surmising that the greatest risk facing Ruby teams’ continued use of Ruby was not only whether Ruby was fast enough or concurrent enough, but whether their existing applications were maintainable enough. The most common reason I’ve seen teams leave Ruby for has been the lack of tools and education needed to work effectively with large, complex applications. (New languages are appealing, but new languages with which you don’t yet associate with tremendous pain are even more appealing!)

Regardless of language, when saddled with a hard-to-change codebase, the grass of a total rewrite will seem greener. But Ruby’s dynamic nature gives us fewer tools than some other language ecosystems for successfully managing the invetible accretion of incidental complexity that every long-lived system faces. Is there anything we can do to make legacy Ruby more maintainable?

That question led me to this talk. In it, I introduces a new gem that we designed to help wrangle legacy refactors. It’s called Suture, and along with providing some interesting functionality to make refactoring less mysterious and scary, it also prescribes a clear, careful, and repeatable workflow for increasing our confidence when changing legacy code.

We hope you’ll watch the talk and try out Suture in your projects! Please tweet at us or open an issue if you have any questions.

00:00
This talk is called surgical refactors
00:03
Could we bring the lights down so that we
00:06
can see the screen better? All right, cool!
00:08
Good morning, let's start.
00:10
So, my name is Searls. That's me on
00:15
Twitter, @searls. That's what I look like
00:17
on Twitter, if you've maybe seen my face
00:20
before. If you want to send me a message
00:23
at justin@testdouble.com.
00:25
My last name, I katakana-ize usually as
00:30
サールズ (saaruzu) if you want to call me
00:32
that, but the pronunciation is difficult,
00:34
so in Japanese you may call me Juice.
00:39
Yes, like "Apple juice".
00:41
I come from a company called Test Double
00:45
We're software consultants. If your team
00:48
is looking for additional developers that
00:51
are really good at Ruby, JavaScript, and
00:53
testing, and refactoring, we would love
00:56
to work with you. And if you're okay with
00:59
people who mostly speak English you can
01:01
send us an e-mail at hello@testdouble.com
01:05
So, first, a little bit about me. This is
01:08
me eating some Yaboton misokatsu in Nagoya
01:12
last week. I talk very fast when I'm
01:15
nervous. And I'm always nervous, so please
01:22
forgive me. Everyone, I'm sorry. Please
01:24
try [to keep up].
01:25
We have plenty of time today, so if I'm
01:31
speaking too fast, it's okay to shout
01:34
"please go more slowly". That sort of
01:37
thing is fine. Speaking of me being
01:41
nervous, I was very nervous about screen
01:43
size with the projector here this morning
01:46
I didn't know if it was 16:9, or 4:3. But
01:49
I think we landed okay, but it actually
01:52
gave me an idea because Matz yesterday
01:55
was talking about Ruby 3x3 and how
01:58
difficult it will be to solve so I think
02:00
can solve it here today really easily
02:02
here it is [a 3:3 aspect ratio Ruby].
02:06
That was a lot easier than he made it
02:07
sound. Speaking of Ruby it's a massively
02:10
successful language, right? And in the
02:13
early days, success looks like a lot of
02:16
happy people, they're building stuff for
02:19
fun. People give us a lot of positive
02:22
attention. I remember when I started
02:24
learning Ruby, everyone I knew who wrote
02:26
Ruby was really cool. And in the early
02:30
goings of any language, the marker for
02:32
success is if you're able to make it easy
02:35
to make new things. And one of the best
02:38
things about Ruby is that building
02:40
something new is very easy. It does that
02:42
very well. But later success is very
02:45
different, because people are more
02:47
critical. Now that it's popular, it's an
02:51
incumbent, people are more critical in how
02:53
they analyze the language. They're more
02:55
serious because people use it for work
02:57
And as time goes on, more and more money
03:01
is involved in the existence of the
03:03
programming language. And so the mood is
03:05
very different with an older programming
03:07
language. And I think the key to later
03:09
success for a programming language is
03:11
making it easy to maintain old things
03:14
And nobody likes maintaining old things
03:16
But my question for today: can we make
03:19
it easier to maintain old Ruby code? And
03:23
so I sat and I thought about this for a
03:26
while. So, today as an exercise, we're
03:29
going to refactor some legacy code. The
03:31
word refactor stands out, you might define
03:35
the word refactor—if you're not familiar—
03:38
is to change the design of code without
03:41
changing its observable behavior. So we
03:43
change the internal implementation, but
03:45
it still behaves the same way. How I think
03:48
of refactoring is to change in advance
03:51
of a new feature or a bug fix, in order to
03:54
make that feature or that bug fix easier
03:56
later. Legacy code is the other phrase in
04:01
this sentence and legacy code has many
04:04
definitions. Here are a few legacy code
04:06
definitions. First: old code. Or some
04:11
people say, legacy code is code without
04:14
tests. Usually, we say legacy code just
04:16
to mean a pejorative. It's code we don't
04:19
like. But today I'm going to use a
04:22
different definition of legacy code. I'm
04:24
going to say that legacy code is code that
04:26
we don't understand well enough to change
04:29
confidently. So today, let's refactor some
04:32
legacy code. Now, whenever anyone says
04:36
"refactor" and "legacy code" in the same
04:38
sentence, I feel like refactoring is
04:40
really hard, because you have to take
04:42
something that exists and then make the
04:44
design better and that just takes a
04:46
certain amount of creativity I don't
04:48
always have. But refactoring legacy code
04:50
is very hard, because it tends to be very
04:53
complicated. And as a result, it's easy
04:56
to accidentally break existing
04:58
functionality for users, and so it feels
05:01
dangerous. As a result, legacy refactors
05:03
on teams often feel unsafe. They make us
05:07
nervous. Additionally, legacy refactors
05:11
are hard to sell to our managers and to
05:14
businesspeople. If we were to chart
05:17
business priority of our activities as
05:19
developers against the cost and risk of
05:21
what we're doing, in the top-right corner,
05:24
in that quadrant: new feature development
05:26
Obviously high priority, but also very
05:29
expensive. In the top-left you'd put bug
05:32
fixes. High priority, but relatively less
05:35
expensive. In the bottom-left, low
05:38
priority from the perspective of the
05:40
business, but still relatively low cost is
05:41
testing. And what's down here in the
05:44
bottom right? I think refactoring goes
05:46
here. So new features, we don't have to
05:49
sell the business on new features. If
05:51
they're paying you a salary, they've
05:52
already decided they want to invest in new
05:54
features. Bug fixes are normally easy to
05:57
sell, because the cost-to-benefit is good
05:59
Sometimes, we succeed at selling testing,
06:02
we usually do nowadays. But we don't
06:04
always, because it's not always seen as
06:05
very important. Selling refactoring is
06:08
very hard, typically in my experience.
06:10
In general, refactoring is just difficult
06:14
It's difficult to estimate it's going to
06:17
take because you don't know how much work
06:19
it's going to be or how risky it is. From
06:21
the business's perspective, it's invisible
06:23
right? If to refactor code is to change
06:25
its implementation, and not its observable
06:27
behavior, they don't know what value
06:30
they're getting out of the refactor. And
06:32
typically, because we're changing
06:34
something that's very complex, we have to
06:36
put a stop on all other work on that area
06:38
of the code because it would be difficult
06:39
to merge in multiple changes, so we're
06:42
stopping everything now. As complexity
06:46
of that legacy code increases, it probably
06:49
means that it was more important, right?
06:51
The business needs something about that
06:53
code to have 500 "if" statements and all
06:56
sorts of complexity, so it's probably a
06:58
very important piece of code. And so
07:00
therefore changing it is less certain. And
07:03
in general, more costly. So today as part
07:08
of my "Make Ruby Great Again" series of
07:10
talks, I want to make refactors great
07:12
again. Of course I thought about that line
07:15
for two seconds before I realized
07:17
refactors have never been great, and so I
07:19
just want to make refactors great for the
07:21
first time, is my goal today. So in
07:25
looking at this quadrant and looking at
07:27
refactoring, there's really two ways we
07:29
could make refactoring easier. On this
07:31
axis, on business priority, we could try
07:32
to sell refactoring to businesses, so they
07:35
view it as a higher priority. Now when we
07:38
sell refactoring to businesspeople, the
07:41
image in their mind is typically like
07:43
road construction. We're going to stop
07:45
everything, no traffic is going to get
07:47
through, but money is going to continue to
07:49
fly out the window at the same velocity
07:51
that it normally does. It's not a very
07:54
attractive image to our managers. So we
07:58
have a few tricks that we use to sell
07:59
refactoring. The first one is that we try
08:01
to scare them into it. We say, "hey, if
08:04
you don't let me refactor this right now,
08:06
someday we'll need to rewrite everything"
08:08
But that's far in the future, that's
08:10
difficult to prove. We might say your
08:13
maintenance costs in the future will be
08:15
much higher, but that's difficult to
08:17
quantify. It doesn't feel real.
08:19
Secondarily, we might try to absorb the
08:23
cost, through discipline and
08:25
professionalism. Maybe these are our new
08:27
feature activities normally. We plan. We
08:30
develop. We write tests. Maybe we just
08:33
grow the pie and spend a little bit more
08:36
time on each new feature by baking in some
08:37
time for refactoring. And this is
08:40
fantastic, but it requires immense amounts
08:42
of discipline that most people don't have
08:43
And, as soon as there's any time pressure,
08:47
refactoring is going to be the first
08:48
practice that we drop. And most teams
08:51
experience a lot of time pressure, so
08:52
that's not very effective either. Another
08:55
strategy teams use is to take hostages
08:58
The business sets the backlog priority
09:00
saying I want feature 1, 2, 3, 4. But the
09:04
team says, "No no, we're going to do some
09:07
refactoring after feature 1 and before you
09:09
get feature 3, we're going to do some more
09:11
refactoring. And this is problematic
09:13
because it's adversarial. Basically, we're
09:16
blaming the business for rushing us and
09:18
telling them that "no, we need to go
09:20
slower. We need to do 'our stuff' now"
09:21
It erodes the business's trust in the team
09:24
because they're paying us a lot of money
09:26
to write code and if the message that
09:29
we're sending them is that we're so bad at
09:31
it that we have to stop and fix it every
09:32
now and then, they're going to think that
09:34
we're less competent at our jobs. So yeah,
09:38
refactoring is hard to sell. We're all
09:40
programmers, we all believe in it, but
09:42
we're probably not going to successfully
09:44
change the culture today. So that's
09:48
probably not where the solution is in the
09:50
short-term. If we look at the other axis,
09:53
why is refactoring so expensive? Well,
09:56
whenever I refactor code I feel a lot of
09:59
pressure. There's a lot that I have to
10:01
keep in my head at once. I have to get
10:04
a lot done, but in a short amount of time
10:06
because other people are waiting on me to
10:09
maybe merge in my changes; because it's
10:12
low-priority so it doesn't get afforded
10:14
very much time to work on it. And my tools
10:18
in general, we don't have a lot of tools
10:20
that help us refactor code. Most open
10:22
source tooling is about creating new stuff
10:24
because that's more exciting and that's
10:25
what we want to think programming is, but
10:27
most of us are getting paid to maintain
10:29
old code, so you'd think we'd have better
10:31
refactoring tools, and we just don't. So
10:35
for me refactoring is very scary, and I'm
10:38
on a mission to find everything that's
10:40
scary about software and try to find a way
10:42
to make it less scary, because I'm so
10:44
anxious and scared all the time. In fact,
10:47
if you're like me and you're scared all
10:49
the time you should buy my book that I'm
10:51
working on, it's called "The Frightened
10:53
Programmer". It's not a real book, because
10:56
I'm too afraid to write a book. So what
11:01
can we do to make refactoring less
11:03
expensive? Well, we do a few things
11:06
already. Like this book by Martin Fowler,
11:08
Jay Fields. And they're explicit
11:13
operations like "extract method" or "pull
11:16
up", "push down", or "split loop". They
11:19
have names, because if you follow the
11:21
procedure carefully enough, it's safe to
11:24
undergo certain refactoring operations
11:27
And it's safer still if you have good
11:29
tools. Easily my favorite thing about
11:32
Java programming language is that with
11:34
Java, it's so not-expressive, that you're
11:36
able to get all these automated
11:38
refactoring tools in your IDE with a
11:40
relative guarantee that you're not going
11:42
to break anything. But they're also not
11:46
very expressive of operations, either. You
11:48
couldn't possibly take a complex design
11:51
and then make it into a good design if
11:53
you only follow that advice. If you only
11:55
follow those operations. Second, a
11:58
technique called characterization testing
12:00
was made popular in the book "Working
12:02
Effectively with Legacy Code" by Michael
12:03
Feathers. And that's still, I think, the
12:05
best advice a lot of people have about
12:08
legacy rescue. Basically, you treat the
12:12
code as a black box. And then you put a
12:15
test harness around it, and you send it
12:17
some input and you get back an output.
12:18
You send an input. You get back an output.
12:20
You send all the inputs that you can
12:22
imagine into the black box and you just
12:24
record the output no matter what it is
12:26
without understanding, without judgment
12:28
There's not wrong answers. The goal is
12:30
simply to create a harness by which you're
12:34
pretty sure that if you change stuff, as
12:36
long as the test passes, the change was
12:38
safe. Once you zoom in and you have that
12:42
test harness, then you can begin to
12:43
aggressively refactor the code into new
12:45
units and new objects and then when you're
12:48
done with that, you can backfill in new
12:50
unit tests that can understand what the
12:52
code is doing and that can have clear
12:53
intention behind them. And that's a lot
12:56
of testing, right? We have to write all
12:58
these characterization tests, that takes
13:00
a lot of time; we have to write new units
13:02
and then we have to write tests for those,
13:03
that's a lot of work. And the next step
13:05
is actually to delete a test. After we're
13:07
done with the refactor, we're supposed to
13:09
delete our characterization tests, but if
13:11
you have a lot of legacy code in your
13:14
codebase, you probably want all the code
13:16
coverage you can get and if you just spent
13:18
a lot of time on a test, the last thing
13:21
you're going to want to do is delete it
13:22
And it's tempting to quit halfway through
13:25
because it's a time-consuming, exhausting
13:27
process, and that's not good. More
13:29
recently, a technique has been popular
13:33
with legacy rescue that resembles A/B
13:36
testing. Basically, if you have the old
13:38
code over here and you write a new
13:40
implementation over here, then you just
13:44
put a router in front of it. Maybe you
13:46
send 20% of the traffic to the new code
13:48
80% to the old code. I think her name is
13:54
Jesse Toth, I think she's here today.
13:56
Jesse, are you here? Yeah! There's Jesse
13:59
Jesse's working on this awesome gem from
14:01
GItHub called Scientist and what it
14:03
requires is you're on-your-own for how you
14:10
rewrite that new code path and a lot of
14:14
sophisticated monitoring and logging and
14:16
data collection that scientist produces is
14:19
necessary to figure out whether or not the
14:21
changes are safe. And finally, it's only
14:24
really appropriate for business domains
14:26
where it's safe for transactions to fail
14:28
It might work for GitHub, but it may not
14:31
work for a bank that's handling financial
14:34
transactions. So if you think of it as a
14:37
spectrum on one end with characterization
14:40
testing on one end and scientist on the
14:42
other end, you can look at Michael
14:45
Feathers' book and say that's good for
14:46
development, it's kind of painful for
14:48
testing, and it has no solution for
14:50
staging or production environments. If you
14:53
look at Scientist on the other side, it
14:56
doesn't have a lot to say about
14:57
development or local testing, it's really
15:00
obviously beneficial for a staging
15:02
environment, and if anything in production
15:05
the amount of data that it produces is
15:07
overwhelming, it's really cool. But what
15:09
if we had a tool that was actually good at
15:13
all four stages of the life of a refactor
15:16
from planning to completion and what would
15:20
that look like? And so that was the
15:22
question that I posed to myself when I
15:24
submitted to speak at this conference
15:27
And then months passed by and I still had
15:29
no good answer and eventually I said "Oh
15:31
no, I have to give a talk on this" and
15:32
being frightened, like usual, I thought
15:37
about it and thought about it and I had an
15:40
idea and I decided that instead of writing
15:43
a lot of slides today explaining how to
15:45
refactor well, I'll just write a new gem,
15:47
because even though I speak English and
15:50
my English may be hard to understand, our
15:53
common language is Ruby and so let's talk
15:55
about Ruby for the rest of the talk. I
15:58
used TDD, so this is a TDD talk. Based on
16:04
that explanation, TDD stands for
16:05
"Talk-Driven Development"
16:10
I like this practice, I'm going to try to
16:11
rely on it more. The tool that I wrote is
16:15
a new gem called Suture. Sutures are the
16:17
stitches that you make when you're closing
16:20
a wound after a surgery, so I think it's
16:24
a good image for surgical refactoring,
16:27
it's up on our GitHub at @testdouble, up
16:30
here; the page looks like this. You can
16:34
install the gem, just like any other gem
16:37
You know how to install gems, I'm sure
16:39
The metaphor here is to treat refactors
16:41
like surgeries. Now surgeries, they all
16:45
serve a common purpose—to help us get well
16:48
We want to take a scary thing, and make it
16:50
feel safe. They require careful, up-front
16:53
planning. They require flexible tools
16:57
Because, you can imagine, this refactor is
17:00
going to be loaded with roughly the same
17:02
information and that same information can
17:04
be used to make development, test, staging
17:06
and production all easier. And we want to
17:09
follow a consistent process, because there
17:11
are already more than enough variables in
17:13
all of our legacy code, because it can
17:15
look all sorts of different ways. So if we
17:17
make the process consistent, then we can
17:19
feel a little better. And we take multiple
17:22
observations. Initially, we need to be
17:25
very up-close to our refactor, but after
17:28
we've initially developed it, we can step
17:29
further back and eventually just assess
17:31
based on logging the health of the
17:33
refactor before we've decided that it's
17:35
complete. So we're going to talk about 9
17:38
features or workflow steps that are
17:41
built into Suture. First, how to plan.
17:43
Second, how to identify a seam that we're
17:47
going to cut. Third, how to record the
17:49
interactions of the old code path and then
17:54
how to automatically in a test environment
17:57
validate that we're able to reproduce all
17:59
of those recordings. Finally, we get to
18:01
refactor the code (or reimplement it).
18:04
Then we verify that the new refactor
18:08
behaves the same way against the same
18:10
recordings. In a staging environment, we
18:12
can compare that both the old and new code
18:14
paths behave the same way, even if they're
18:18
getting input that's different from what
18:19
we've gotten before. And in production, we
18:21
can use the same information as a fallback
18:24
mechanism to rescue any errors in the new
18:26
code that we didn't otherwise anticipate
18:29
Finally, the last step of Suture is to
18:32
delete Suture. Just like you have stitches
18:35
removed, Suture shouldn't be in your
18:38
Gemfile forever, only when you're doing a
18:40
legacy rescue.
18:41
So, a little about planing. Today in this
18:44
exercise, we're going to do two bug fixes
18:47
The first bug fix is we have a calculator
18:50
service and this calculator service has an
18:53
add route but it doesn't add negative
18:56
numbers correctly. It's a pure function
18:59
so if you look at this controller method,
19:01
we create a calculator, we call add with
19:04
two operands, and then we set it to result
19:06
Just like a Rails controller method
19:09
If you look at the implementation, you can
19:11
see we've got the declaration of the
19:14
function and then for whatever number of
19:17
times, the right operand is, like 8 times
19:20
it'll loop over 8 times add 1 to the left
19:24
and return the left, now that's where the
19:27
bug is obviously because it's only
19:29
positive. And yeah, this legacy code is
19:31
really ugly, but I'm sure your legacy code
19:34
is really ugly, too, so that's the idea
19:38
So if we zoom back out, we're going to
19:41
create our seam here. This is going to be
19:44
the point at which we want to branch
19:46
between the new code and the old code
19:48
because it's the call-site, it's the most
19:50
common-sense place for us to cut it. The
19:53
next one, from a planning perspective, the
19:55
next bug is: we have a tally service where
19:59
we can invoke the calculator multiple
20:03
times and then ask it what the total sum
20:04
of all the numbers that we called it with
20:05
was. And this is a different type of
20:07
function, because it's going to require
20:09
mutation of an object over time. Here's
20:11
what that method looks like; we again
20:13
create a calculator and then over an array
20:16
of numbers, for each of them we call a
20:18
tally function. Finally, we take the total
20:20
from the calculator and set it to the
20:22
result. When we zoom in here, this is an
20:25
even more ridiculous-looking function, we
20:28
instantiate a @total instance variable,
20:31
and then we count down from whatever was
20:34
passed to 0, and if it happens to equal
20:37
exactly half of what's sent, then we add
20:40
double it, and this is really ugly, but
20:44
it was a way to introduce the bug, right?
20:46
So this will only work for even numbers,
20:48
it'll have no effect if we pass it an odd
20:50
number and we want to fix that, that's
20:52
our story today, to fix that bug. We zoom
20:54
back out, obviously that's our call-site
20:58
just like before, but it's bigger than
20:59
that, because we also depend on the value
21:03
of that @total instance variable, and so
21:06
this seam is going to be more complex
21:07
It's going to require a little bit more
21:09
work. Next, let's talk about how to cut
21:13
those seams. So in the pure function case,
21:17
what we do is we use Suture's API, so
21:20
zooming out a little bit, instead of
21:22
calling calc.add directly, we're going to
21:23
create a Suture. We're going to say
21:25
Suture.create a seam called :add, and
21:28
then we're going to pass it the same
21:29
arguments as an args option, as an array,
21:32
and we're going to pass it a reference to
21:34
the callable add method. Now, this could
21:37
be any callable, it could be a lambda, a
21:39
Proc, it could be a reference to a method
21:41
like this, it doesn't matter as long as it
21:43
can be called with `call`
21:45
And at first, what I've just done here is
21:48
a no-op; I'm going to write exactly this
21:50
much and then go to the page and make sure
21:52
that I did this correctly, but it's not
21:54
going to take any further action, it's
21:56
just going to call through like it
21:57
normally would. In the mutation case,
21:59
this is going to be more complicated. So
22:02
let's zoom out again, give ourselves some
22:04
space to work and I'm going to change that
22:07
tally call to another Suture, creating
22:09
:tally and the args here, I'm going to
22:12
say are the calculator and the number, and
22:14
you might say to yourself, "wait, tally
22:17
just takes this one argument, how is that—
22:18
—what's going on?" Well, to design a seam,
22:22
we need to think in terms of the impact of
22:23
the function. Pure functions are really
22:26
easy, right? We treat them like a black
22:28
box, we pass in a couple of arguments, we
22:31
get a return value. We pass in the same
22:34
couple of arguments, we know we're going
22:36
to get the same value. They're repeatable
22:38
input and output. And recording a bunch of
22:40
those is going to be safe and make sense
22:42
Mutation is a lot more difficult. In the
22:46
case of this tally function, if I pass it
22:48
4, I get 4. But if I pass it 4 again, I'll
22:50
get 8. And so if you think more broadly,
22:53
it's because this instance variable is
22:56
changing and so the real argument of this
22:59
is two-fold. First, there's the calculator
23:03
and what's the state of the calculator and
23:05
then there's the number we're passing to
23:07
it. And so if we fix both of those
23:10
invocations to pass in the calculator at 0
23:12
then we'll actually get back to a
23:15
repeatable input and output. So if you
23:16
think of it that way, we're just
23:20
broadening the seam of the cut that we're
23:21
going to make and here we're just going to
23:23
pass an anonymous lambda, and in that
23:26
lambda we're going to call tally on the
23:28
calculator that gets passed in for the
23:30
number that gets passed in , and then also
23:32
we're explicitly returning the total so
23:34
that we get a clear input and a clear
23:36
output. That'll make our recordings more
23:38
valuable. Once we've recorded we want to
23:43
make sure—oh, excuse me, got ahead of
23:45
myself—now we've got to record the
23:47
interactions at the seams that we just
23:51
made. So, in the pure function's case, all
23:54
we have to do is add this option that says
23:57
record_calls: true and because in legacy
24:00
environments, often-times we might want to
24:02
deploy code and not actually make code
24:04
changes to our Ruby files, almost all of
24:06
Suture's options can also be set with
24:07
environment variables like this. And to
24:11
record some calls, all we have to do is
24:13
call the thing. You could use the command
24:15
line. You could just set some params,
24:17
call show, that's going to record an
24:19
invocation. Set different params, call
24:21
show, that'll record. And so on and so
24:24
forth. You can also record via the browser
24:27
if you've got a route, you can just go to
24:29
the browser and keep refreshing the page,
24:31
those will record as well. You can even,
24:33
if you need to, you can record in a
24:35
deployed environment like production and
24:37
pull down both Suture's snapshot and a
24:40
production snapshot to replay off of that
24:43
for cases where you don't have a good
24:45
solid reliable development database. Now
24:49
in the case of the mutation, we're also
24:52
going to set record_calls: true and
24:54
because we already did the hard work of
24:55
identifying a seam, it'll work the same
24:57
way. We can set some parameters and call
24:59
index. That'll record. We can set
25:02
different parameters, call index, and
25:03
repetitively do this to generate some
25:06
recordings. Now, where do those recordings
25:08
go? Well, by default, Suture is going to
25:11
assume you have a database, a SQLite
25:13
database, at db/suture.sqlite3 and you
25:17
don't have to set this up. This is
25:19
invisible to you. You'll just notice that
25:21
a database shows up. You don't have to
25:23
worry about the schema or calling it, it's
25:24
just a place to save stuff. I did this
25:27
because I heard Ruby was getting a
25:28
database yesterday. Matz was talking about
25:31
Ruby 3 and databases so I figured that
25:34
was a good approach. All it really does
25:37
is dumps—using Marshal.dump—the inputs and
25:40
the outputs so we can record the calls.
25:42
So you might ask yourself, "Does this
25:44
scale? Does this work with Rails?" I was
25:46
also curious about this, because I spent
25:49
100 hours on this gem before checking to
25:50
see if I could use it with Rails. I'm
25:54
going to look at an example, the Gilded
25:55
Rose Kata, this is an exercise for
25:58
practicing refactorings, it's up on Jim
26:01
Weirich's GitHub, here. And here, I
26:05
intentionally put it into a Rails
26:08
controller, and you don't have to read it
26:10
too closely other than to see this is a
26:12
Suture that takes in items and returns
26:13
items after they're updated. So I'm going
26:17
to go to the page. This is the little page
26:19
that I made. This repo, by the way, is
26:21
inside of Suture's repository, and you're
26:23
welcome to pull it down and play with it
26:24
And obviously this is a beautiful web site
26:27
but it's just there as an example. I'm
26:29
going to create a few items. So here's a
26:32
normal item, and another item, and a third
26:34
item, and I'm going to put them in a table
26:38
and then hit this Update Quality button,
26:40
because that's going to invoke the
26:42
function that I want to record. And you
26:45
can see that the table changes, and if you
26:48
look at the SQLite database, there's now
26:51
a few invocations that have been recorded
26:53
And I'll click the button again. And a
26:54
couple more invocations. And if I click
26:56
the button again, a couple more. I keep
26:58
doing this to generate all of the
27:00
different cases I want to get better test
27:02
coverage. So yeah, Rails apparently works
27:04
So that's cool. That's reassuring, because
27:06
there's a lot of us who have legacy code
27:08
have legacy Rails. Next, we have to take
27:13
those recordings and validate that we can
27:15
reproduce them. Otherwise, they won't have
27:16
any value to us later. So in the case of
27:18
the pure function, we're just going to
27:19
write a little test. We're going to test
27:21
that if we create a calculator, and then
27:24
we call Suture.verify with the same name
27:26
that we just recorded everything as, and
27:29
then we pass it the subject method, which
27:31
is the method under test—the old method in
27:33
this case. Once we do that, it's going to
27:36
load the database for that name and then
27:38
go over each recorded call and compare the
27:41
recorded arguments against the return
27:43
value. You can think of it in terms of
27:47
Michael Feathers' book. We just created a
27:49
black box and put some characterization
27:51
tests around it. And you get these tests
27:54
basically, for free. Instead of writing a
27:57
lot of tests, we can just write a few. In
27:58
the case of the mutation, we write a very
28:00
similar-looking test. So here, we pass it
28:02
a lambda, called :tally and we return the
28:06
@total. Note that we want to duplicate the
28:09
lambda exactly, because it needs to behave
28:11
just like the other one in order to work
28:13
with the recordings.
28:14
What's interesting about this is that it's
28:16
actually a really good use for using a
28:18
code coverage tool. So if you look at the
28:21
Gilded Rose Kata, and you look at how Jim
28:24
wrote the initial tests for it, this is
28:26
the initial test file. He had to read all
28:30
the code and then write hundreds of lines
28:32
of tests, understanding all of it. And
28:34
clearly he put a lot of investment into
28:37
writing those test cases. That's a lot for
28:40
a test that you're supposed to delete
28:41
eventually. So, with Suture, you can take
28:44
the exact same codebase and write a little
28:47
test like this one. Pass it a lambda
28:50
Items come in, items come out and Suture's
28:53
smart enough to dump out the arguments at
28:58
invocation-time and then dump them out
29:00
again after invoking the function. Even
29:02
though the items are just mutated here,
29:05
these calls work fine. Additionally, we
29:08
have a fail_fast option, normally it
29:10
aggregates all your errors to give you
29:11
good messages, but here if you expect all
29:15
the recordings to pass, it'll fail fast,
29:17
which is a little bit more performant,
29:19
especially if you've got a lot of database
29:20
interaction.
29:22
Before we call it finished, we can check
29:25
code coverage. If you look at the code
29:27
coverage here, you can see that it's all
29:28
green, but if there was a particular case
29:30
that was yellow or red, I'd be able to
29:32
read that, and then all I'd have to do to
29:34
cover that with a test is to create an
29:36
item that matched those conditions and
29:39
record it. You can get 100% code coverage
29:42
writing no tests at all, which is really
29:44
pretty cool. Now comes the most important
29:48
step, the center of the square. How to
29:51
refactor the code. Unfortunately, I'm
29:53
really bad at refactoring. I don't know
29:55
anything about refactoring, really, it's
29:57
not my forte. That's why I needed this
30:00
tool, to make it more safe for me. If you
30:02
want to learn about refactoring, my
30:04
friends Sandi and Katrina wrote a book
30:06
this year called "99 Bottles of Object-
30:08
Oriented Programming" and it focuses a lot
30:11
on refactoring techniques for Ruby. Unlike
30:14
my book, this book actually exists! You
30:18
can go buy it and enjoy it. In the case
30:23
of the pure function, how we can refactor
30:25
this—it's a simple method, right? It's
30:26
adding. If we remind ourselves of what the
30:30
bug was, which is that it doesn't work
30:32
with negative values. Then we can create a
30:34
new function and do the simple thing,
30:36
right? left + right, that's all it really
30:38
needs to do. We're going to do something
30:40
else, as well. We're going to actually
30:42
reproduce the bug. So we're going to
30:44
return left if it's negative. We want to
30:48
retain all of the behavior that it
30:51
currently has, so it passes against all
30:53
the recordings and the real reason we do
30:55
that is we want to just change one thing
30:57
at a time. We're here to refactor the code
30:59
to prepare ourselves to fix the bug. We're
31:02
not here to fix the bug yet. It would be
31:05
arrogant to try to fix it now, because we
31:07
honestly don't know where else in the
31:10
system, it's actually depending on the
31:12
unintended consequence of the bug. Right
31:15
now we just want to make sure we can
31:17
refactor it safely. In the case of the
31:20
mutation function, you can see this is
31:23
just a total mess, and remember that the
31:25
bug is that it skips all of the odd values
31:27
So we have to be thinking about that
31:29
We're going to create a new tally function
31:32
Here we instantiate the ivar, we add the
31:35
number, and then we return nil, just
31:38
because the other one returned nil, and
31:40
we want it to behave as identically as
31:42
possible. But we want to reproduce the bug
31:45
so if it's an odd number, we'll just
31:48
return as a short-circuit up at the top,
31:50
with a FIXME to remind us that this is
31:52
what we're here to change later. Kent Beck
31:54
said, "make the change easy, then make the
31:58
easy change." So that's the mindset of
32:01
this kind of refactor step. And that's all
32:03
we really have to do to refactor the code
32:05
in this case. Verifying is interesting,
32:08
because now we're going to use the same
32:09
recordings against the new code. And in
32:12
the case of the pure function, we write
32:15
another test and that test is going to
32:17
look exactly like the first test, except
32:19
instead of calling the :add method, we're
32:21
calling our :new_add method. And it just
32:24
passed right away, because pure functions
32:26
are easy—inputs and outputs. In the case
32:28
of the mutation step, we have to again
32:32
replicate the same kind of test, so call
32:35
Suture.verify, we pass it the same-looking
32:37
lambda and it fails. It fails because our
32:42
refactor, actually, there was a bug in it
32:43
So what's going on here? Well, speaking of
32:47
I mentioned Jim Weirich earlier, before
32:49
Jim passed, one of the things he taught me
32:51
is that any library is only as good as its
32:55
error messages are helpful, so I wanted to
32:58
write very helpful error messages. So this
33:01
is an example of one of Suture's error
33:02
messages. It's going to, whenever a
33:05
verification fails, print out a customized
33:07
README to help you fix the bug that's in
33:10
your code, including your progress to date
33:14
If we look at it and we zoom in, the first
33:16
thing it does is it lists out all of the
33:18
failures, just like a test runner would
33:20
and you can see here that it's got ideas
33:24
to fix that thing underneath each one
33:26
In the first case, there's a focus mode so
33:29
you can just focus on one thing by setting
33:30
this environment variable to the ID of
33:32
that recording. Additionally, if you think
33:34
that a particular recording was made in
33:37
error, you can delete it from a console
33:39
I wanted to give good advice about solving
33:44
these failures, so there's advice at the
33:46
bottom of the failures. The first one
33:49
suggests that maybe you want to build a
33:51
custom comparator to compare whether the
33:53
old thing and the new thing's return
33:55
values are comparable. To compare results
33:58
by default, it's very simple. Suture just
34:01
assumes that the thing on the left will
34:03
pass a double-equals test against the
34:05
thing on the right. Alternatively, if that
34:08
fails, it'll also consider things
34:10
equivalent if they Marshal.dump to the
34:12
same string value. You might ask about
34:15
ActiveRecord. It has some custom stuff
34:18
in place in case that it detects that the
34:20
two objects are ActiveRecord and
34:22
essentially what it does is it compares
34:23
their attribute hashes. There's some more
34:25
to this and there's an option that you can
34:27
exclude certain attributes if you want to
34:29
I think by default it just skips
34:32
created_at and updated_at
34:35
But what if, even still, your two things
34:36
don't equal each other? Well, that's not
34:39
good, so I needed to create a way for you
34:41
to have custom comparators. Here you can
34:44
define a comparison however you like. So
34:45
if you look at our calculator example
34:47
Just pretend it had a lot of other fields
34:50
that weren't very important for the
34:51
purpose of our refactor. What we could do
34:54
in the case of, say, maybe we returned
34:56
calculator instead of just the total
34:58
number, we could write a custom comparator
35:00
that would take the recorded and the
35:01
actual values and we would just simply,
35:03
just as if you were writing a custom
35:05
comparator in any other Ruby, or a custom
35:08
equals method, just compare the total
35:10
value between the two in order to get to
35:12
a passing, working state. Classes are also
35:16
a thing, so you don't have to use so many
35:18
anonymous lambdas, you can also create
35:20
your own class. You can even extend our
35:22
default comparator. So you could call
35:25
super and if it passes that, that's fine,
35:27
otherwise call some custom logic. And then
35:30
all you do is pass an instance of your
35:31
comparator into the Suture. So going back
35:35
to the error message, there's a little bit
35:38
more here. All of the tests are run in
35:39
random order by default, and it's got a
35:41
generated seed, so the error will tell
35:43
you what seed it ran at, just in case you
35:45
come across something that was a bug that
35:48
only happens in a particular ordering, you
35:50
can lock it down. If you know you're in a
35:53
situation where insertion order has to be
35:55
the order you call everything in, you can
35:57
also turn off the randomness by setting
36:00
this to nil. And there's other
36:02
configuration options too, and I wanted to
36:04
make those discoverable. So the error will
36:06
list off how this particular verification
36:09
was configured. It'll tell you the
36:10
comparator you're using, where the
36:12
database is, whether you're failing fast
36:14
you can limit things like how many calls
36:16
you test, because maybe it's really slow
36:18
or set an explicit time limit to only
36:21
budget maybe 5 minutes in your build. You
36:25
can also limit how many error messages get
36:27
printed, because you don't need to see a
36:28
thousand error messages typically and it
36:30
will tell you what the random seed was.
36:32
Finally, I wanted to give a sense of
36:35
progress, because if you're refactoring
36:36
something you expect it to always pass,
36:38
but if you're going to reimplement
36:39
something, you start from zero and slowly
36:41
build up. If you look at the bottom, the
36:43
result summary tells you how many passed
36:45
how many failed, the number of total calls
36:47
And a little progress bar at the bottom
36:49
Selfishly, the reason I did this is that
36:52
I wrote a progress bar gem 5 years ago and
36:54
I never had a chance to use it, so this is
36:56
the first thing that uses my progress bar
36:58
gem and that made me pretty happy
37:00
The idea is to give yourself a sense of
37:03
progress over time. Again, I think it's
37:06
really important and a takeaway from this
37:09
talk to think about the messages that your
37:10
gem produces. So all that to say, remember
37:14
our test failed, right? Why did it fail?
37:17
Let's look at the message now. In this
37:19
case you can see expected returned value
37:22
of 0, actual returned value was nil. And
37:26
if you look at the calculator, it's state
37:27
was nil as well. And this is the only case
37:29
that failed. Let's look at our code. And
37:31
what we realize looking at the code is
37:33
that we're too aggressively returning if
37:35
something's odd. In the case that we only
37:38
call with odd stuff, @total never gets set
37:40
so it's nil. The solution is simply to
37:42
move this up to the top of the method and
37:44
if we do that, the test passes. So in that
37:47
case, the test was already a little bit
37:49
useful to help guard against this refactor
37:52
Once we've verified that the new code path
37:56
behaves the same as the old code path with
37:59
the data available to us, an interesting
38:02
thought is what—in English we might call
38:06
this "double-entry accounting"—you do it
38:09
over in the new code path and then you do
38:11
it over in the old code path to make sure
38:13
that it's consistent. Remember, our
38:15
mission here is to make development happy
38:18
testing happy, staging happy, and also
38:20
production happy. And our progress so far
38:23
is that we've only really concerned
38:25
ourselves with development, testing and
38:27
we haven't yet addressed staging or
38:29
production. If we look in the case of the
38:31
pure function, all we have to do is add a
38:33
little parameter called call_both and
38:35
it'll call the new code, and then it'll
38:38
call the old code and make sure that they
38:40
return the same return value, otherwise
38:42
it'll raise an error, which is usually
38:44
safe to do in a staging environment. In
38:47
this case, it just works. But you'll find
38:50
that in practice, in staging and
38:52
production environments, you tend to get
38:54
a lot more interesting data than you can
38:56
generate locally, so my hope is that this
38:59
feature will be valuable to people who are
39:01
trying to get something through a rigorous
39:03
QA exercise. If we look at the mutation
39:06
function, here we just set the same thing
39:10
call_both, but unfortunately it doesn't
39:13
work right away. We get another huge error
39:15
message. This is a MismatchError, and if
39:19
we zoom in, it's telling us that the
39:22
calculator with total 2 and argument 2,
39:24
the new code path returned 2 here, but
39:27
the old code path returned 4. Why is that?
39:31
If you look at it, it's because we're
39:34
passing the calculator in and then
39:36
changing the calculator, so there's an
39:38
additional option when you know that
39:39
you're mutating the arguments where you
39:41
can dupe your arguments—clone them—before
39:43
each invocation. This protects against arg
39:46
mutation. So I set that and I think "OK,
39:48
cool, this will work." But unfortunately,
39:50
it still doesn't work, and the reason is
39:53
now calc is never getting mutated, because
39:56
we're just duping everything before every
39:57
call, so we actually have to update our
40:00
lambda or else total will always be nil
40:02
So we zoom out and now we have to
40:04
reassign calc inside of the lambda in each
40:07
of these cases, and we get back to a
40:09
working state. Now, this is a little ugly
40:12
This is a lot of boiler-plate code
40:15
configuring this thing. You have to
40:17
remember, each of these modes is optional
40:19
Maybe you only use Suture for one of its
40:21
four uses, but keep in mind too that if
40:24
you're genuinely dealing with some very
40:26
nasty legacy code, you have to think of
40:28
the trade-off of do you want to go slow-
40:31
and-safe or is it okay to go a little bit
40:34
faster and potentially make more errors?
40:37
And that's a judgment call that you have
40:38
to make. So that's what it's like in
40:41
staging. In production, I want to be able
40:44
to fall back. One of the reasons that
40:47
refactoring is so unsafe-feeling is that
40:50
I empathize with users. I don't want to
40:53
make a change that's going to produce a
40:54
bad result for people using my software
40:56
What Suture does, is if the new path
41:00
errors out, then you can just try the old
41:03
one; it'll rescue to the old one. In the
41:05
case of the pure function, I change
41:07
call_both here because now I'm going into
41:08
production, I'll set fallback_on_error to
41:11
true. If :new ever raises an unexpected
41:13
exception, it'll just call :old and return
41:16
that instead, and it should be invisible
41:18
to the user. And in this case, because
41:22
this function works, it just works. In the
41:23
mutation case, we already did the hard
41:27
work in the last step to make sure that
41:29
both these things can be called safely, so
41:31
we just change call_both to
41:33
fallback_on_error and that works as well
41:35
Obviously, in production environments,
41:38
it's going to be much faster to just call
41:40
the new one most of the time than trying
41:42
to call both every single time. And
41:44
there's also going to be fewer side
41:46
effects. There won't be any additional
41:49
side effects, except for in exceptional
41:51
cases. So, overall, it should be safer.
41:53
If there's ever a rescued error, Suture
41:56
has its own logging system built-in, and
41:58
you can go check in production and see
41:59
that there's not any logs about unexpected
42:01
errors before you ultimately call your
42:04
refactor "complete." Sometimes, code
42:07
raises errors expectedly, so you can
42:10
register certain error types as being
42:12
expected. That way we'll know not to
42:13
rescue them and we'll allow them to be
42:15
propagated normally. That's about the
42:18
production story. Now the last step is the
42:20
most fun, because we get to delete all
42:22
of my code, so just like stitches, we're
42:25
going to remove Suture completely once the
42:29
wound is healed, once we've decided that
42:31
the refactor is complete. In the case of
42:33
the pure function here, we can look at
42:36
this little tiny test and unlike the
42:37
really long test that I showed before, we
42:40
feel completely fine blowing this away
42:41
The next thing we do is we can open up a
42:43
console and delete all of our recordings
42:46
for that particular seam. And those just
42:49
go away. And we look at our controller
42:51
method and there's a lot of cruft here,
42:53
but if you start looking at it more
42:55
closely you can say "I can kill those two
42:58
things, I can remove this option, I can
43:00
remove all the parameters, and then just
43:01
change this to the method and then shrink
43:03
things back down to look normal again and
43:06
now I'm done removing it from the first
43:07
bug." In the case of the second
43:10
function, it's the same thing, we can
43:12
delete the tests safely, we can delete all
43:14
of our recordings safely, and then—you
43:16
know, this is really ugly here—we can
43:18
start deleting all the stuff we wrote here
43:20
eliminate that. Get rid of that whole
43:23
broadening-the-seam wrapping that we did
43:25
Then just return that and shrink it back
43:28
down to a normal-looking function. So now
43:31
We're pretty much done. We did it! We just
43:34
went through a very very safe and rigorous
43:35
approach this software and it was thanks
43:38
to Ruby's dynamic nature that let us do it
43:40
which I thought was pretty fun. Suture is
43:44
ready to use. This is the first time I've
43:47
ever written a gem and then not really
43:50
shared it with anyone or worked with it
43:52
with anyone prior to release, but as part
43:56
of a talk, so I'd love for all of you to
44:00
play with it and test it. The GitHub is
44:02
again, at testdouble/suture. I've also
44:06
declared today to be 1.0.0, so even though
44:08
it has zero users, I'm going to keep this
44:11
API stable going forward. And together if
44:15
we work together on this, tools like
44:17
Scientist, tools like Suture, different
44:20
ways to think about it, we can make
44:21
refactors less scary and I think that as
44:25
part of the overall theme, we may be able
44:27
to make Ruby more palatable for businesses
44:30
to continue using longer-term, because
44:32
we're making it easier to maintain legacy
44:35
code and that's really important over the
44:36
long lifespan of a language. There is one
44:41
last thing I want to share. Just a little
44:43
story about how I met Ruby. A long time
44:47
ago when I was in college, I lived in Shiga
44:49
prefecture, I lived in the city of Hikone
44:53
If you don't know Hikone castle, then you
44:57
might know Hikonyan. Maybe if you're not
45:01
Japanese, you haven't seen this samurai
45:03
cat mascot who's become synonymous with
45:07
the city of Hikone. This is what Hikonyan
45:10
looks like in person. If you go to Hikone,
45:14
Hikonyan is everywhere. He's all over the
45:16
place. I was in a homestay, and my
45:22
homestay brother, he was a programmer, too
45:24
He had a big bookshelf of Japanese
45:25
techbooks and I thought that was pretty
45:28
interesting. One of them was this year
45:31
2000 book, Programming Ruby, and I'd only
45:33
just heard of Ruby, because it was still
45:35
very new in America. It was fun to see, to
45:38
flip through the pages of this really old
45:39
book. Of something that I thought of as
45:41
very new. And I talked to my friends back
45:44
home and said "Hey look, there's this
45:45
Ruby language, I'm having a lot of fun
45:47
trying to learn it in Japanese." And they
45:49
called back and said—this was right when
45:50
Rails was 0.7 or 0.8—yeah, this is going
45:55
to be the next big thing. It was a really
45:57
cool experience to have as an American
45:59
living in Japan. Eventually I went home
46:01
and a lot of time passed. I've been doing
46:04
a lot of Ruby and a lot of Rails for a
46:05
long time now, but I have to say, getting
46:08
to come back today to talk with all of you
46:10
about my experience with Ruby is very
46:13
precious to me, so I thank you very much
46:15
for this opportunity. Everyone, thank you
46:18
for the opportunity to share my story!
46:21
I am overwhelmed with appreciation 💚
46:27
Really, thank you very much!
46:30
Again, my name is @searls on Twitter. I'd
46:31
love if you followed me on Twitter, we
46:33
could become friends. Tell me what you
46:35
thought of the talk. I'm also, my wife & I
46:39
are going to be in Kansai all month. I
46:42
think we go, we leave on October 3rd. So
46:44
if you live in Kansai and you want to go
46:45
for coffee or curry in Kyoto or Osaka, we
46:49
would love to meet you.
46:50
One more time, thank you!
46:53
[Subtitles were sponsored by @testdouble]

Slides

This talk’s slides are available on Speakerdeck:

Here are links to a some of things I referenced in the talk, in roughly the order they appear:

Justin Searls

Person An icon of a human figure Status
Double Agent
Hash An icon of a hash sign Code Name
Agent 002
Location An icon of a map marker Location
Orlando, FL