Home / Work / Jira performance

Jira performance · Atlassian

Made Jira up to twice as fast for its largest boards, and lifted retention 14%

Customers said Jira was slow. The average load time said it was fine. The average was hiding the customers who were suffering.

  • Role Product Manager, Jira Software Technical Performance
  • When July 2020 to September 2021
Chart of Jira board load time falling over a year
1.5–2×Faster boards and backlogsFor the slowest tenth of loads
2×Faster interactionsOver the same year
+14%RetentionFor users with boards of 1,000 to 1,500 tasks

Overview

"Jira is slow" was one of the most common things customers told us

Some enterprise customers with large boards were leaving, and many didn't even call it slowness: they reported bugs. The dashboards disagreed with them. Average load time looked acceptable and wasn't getting worse.

I was the product manager for Jira Software's technical performance, and none of us on the team, me included, had done performance work before. "Make Jira faster" is about as ambiguous a brief as you can get.

Insight

The average was hiding the customers we were about to lose

I changed the measure from the average to the 90th percentile: how slow Jira was for the slowest tenth of loads. Then I built a cohort in SQL of boards with 1,000 to 1,500 tasks. Those boards were the slow ones. They belonged to larger customers, and the people using them were quietly leaving.

Why the average hid the problem
The average: looks fineThe slowest tenthLarge boards, larger customersFast loadsSlow loads
An illustration of the shape, not real data.

Shipped

A year of fast experiments, four of which moved the line most

Chart of time to interactive falling from July to June, with four labelled fixes
Time to interactive over the year. Each drop is a fix shipping.

Defer hidden components

Don't load what the user can't see yet.

Truncate server rendering

Send less up front.

Parallel database queries

Ask for everything at once, not in turn.

Virtualise long lists

Draw only the rows on screen.

How long the slowest loads took, with before set to 100
Before100
After, at 1.5 times faster67
After, at 2 times faster50
1.5 to 2 times faster means the slowest loads took between half and two-thirds as long.

Trade-off

The fastest fix broke a habit, so we asked what the habit was for

Drawing only the rows on screen was one of our most effective techniques. It also broke the browser's own find, Cmd-F, because the browser can only search what has been drawn. People noticed.

Rather than treat it as a performance problem, I worked with engineering and design on what people were actually trying to do. In a large backlog, they weren't searching the page. They were trying to find an issue among thousands. So instead of protecting the browser's find at any cost, we invested in a search that worked across everything, drawn or not, and pointed Cmd-F straight at it. It fit a core, high-traffic job and took a few sprints, against alternatives with lower impact or longer lead times.

Changing a shortcut people use without thinking needs explaining, so a small prompt told them what Cmd-F did now, and why it was better.

A prompt in the Jira backlog: Cmnd+F takes you directly to the search field, just like keyboard shortcut F. Both shortcuts give you quick access to our optimized search.
The prompt that explained the new Cmd-F. The habit stayed. What it opened changed.
Performance work almost always forces trade-offs in the experience. The job is deciding which behaviours are core to what people are trying to do, and which can change.What I took from it

Team

A team new to performance work, and how it ran

Before we touched the backlog, I ran an induction on the problem with performance experts from across Atlassian: what we were solving, why it mattered to customers, and what good looked like. You can't motivate people towards something ambiguous, so the first job was making it concrete.

Leads who co-wrote the plan

The engineering manager nominated two feature leads, one for the backlog and one for the board. They built the strategy with me, and I published it internally with my name on it.

Templates for experiments

Any engineer could run a performance experiment without reinventing the process, which kept the pace up for a year.

Credit by name

Every win was posted with the engineer who made it, so the people doing the work were seen doing it.

A year-end write-up

Published internally and publicly, so everyone who ran an experiment had one place to point to.

By the end, one of the two leads had been promoted to Principal Engineer and the other became an Engineering Manager. I can't claim I caused that. I did make sure their work was visible and linked to them.

Customers

Customers noticed before we told them

About four weeks after the first changes shipped, a customer replied to one of my posts on the Atlassian Community. They had been close to giving up on Jira for planning altogether.

A customer's comment on the Atlassian Community, October 2020: I've noticed a huge speed increase since 4 weeks ago. Managing the backlog during our sprint planning is now really possible! Thanks so much for focussing on this, I was about to start doing sprint planning in excel and loading in a csv to Jira, it was honestly so terrible before. Ivan replies inviting them to the Community Beta Group.
October 2020. The customer we were about to lose, in their own words, and my reply inviting them to try changes early.

I wrote up the results for customers in a post on the Atlassian Community.

Reflection

An average describes nobody

The right metric is the one that shows you the customers you are about to lose. Since then, the first thing I ask of any headline number is who it hides.