Infrequently Noted

Since 2015 I have been lucky to collaborate with more than a hundred teams building PWAs and consult on some of the world's largest sites. Engineers and managers on these teams universally want to deliver great experiences and have many questions about how to approach common challenges. Thankfully, much of what once needed hand-debugging by browser engineers has become automated and self-serve thanks to those collaborations.

Despite advances in browser tooling, automated evaluation, lab tools, guidance, and runtimes, teams I've worked with consistently struggle to deliver minimally acceptable performance with today's popular frameworks. This is not a technical problem per se — it's a management issue, and one that teams can conquer with the right frame of mind and support.

What is Performance? #

It may seem a silly question, but what is performance, exactly?

This is a complex topic, but to borrow from a recent post, web performance expands access to information and services by reducing latency and variance across interactions in a session, with a particular focus on the tail of the distribution (P75+). Performance isn't a binary and there are no silver bullets.

Only teams that master their systems can make intentional trade-offs. Organisations that serve their tools will tread water no matter how advanced their technology, while groups that understand and intentionally manage their systems can succeed on any stack.^[1]

Value Propositions #

The value of performance is deeply understood within a specific community and in teams that have achieved high maturity. But outside those contexts it can be challenging to communicate. One helpful lens is to view the difference between good and bad performance as a gap between expectations and reality.

For executives that value:

• Revenue
  Performance is a significant revenue contributor. To the extent that a product performs poorly, a fraction of revenue will be lost and the narrower the funnel becomes for all types of conversions.^[2]
  
  
• Engagement
  Poor performance has a well-documented relationship to reduced engagement. The space between good and bad performance is the opportunity for a competitor's product to fill the same opening. An under-appreciated aspect of this equation is variance. When a product often performs well but sometimes is slow to respond, users lose trust and use that service less. Consistent performance matters just as much as low average latency.
  
  
• Design
  Poor performance is a gap between the responsiveness of Figma mockups and brand reality. Products that perform well are closer to the approved design. Many brands build their reputations on fanatical attention to detail about products and the physical environments they're sold in. A laggy digital experience is out of alignment with these values.
  
  
• Accessibility
  Performance is the foundation of access. Users experiencing slow services may be unable to access them in the first place, mooting other work poured into improving a11y. Reduced network and device capacity correlate with other access challenges. a11y that isn't founded on consistently excellent performance can easily veer into performative, rather than effective, territory.

Performance is rarely the single determinant of product success, but it can be the margin of victory. Improving latency and reducing variance allows teams to test other product hypotheses with less noise. A senior product leader recently framed a big performance win as creating space that allows us to be fallible in other areas.

Protecting the Commons #

Like accessibility, security, UI coherence, privacy, and testability, performance is an aggregate result. Any single component of a system can regress latency or create variance, which means that like other cross-cutting product properties, performance must be managed as a commons. The approaches that work over time are horizontal, culturally-based, and require continual investment to sustain.

Teams I've consulted with are too often wrenched between celebration over launching "the big rewrite" and the morning-after realisation that the new stack is tanking business metrics.

Now saddled with the excesses of npm, webpack, React, and a thousand promises of "great performance" that were never critically evaluated, it's easy for managers to lose hope. These organisations sometimes spiral into recrimination and mistrust. Where hopes once flourished, the horrors of a Bundle Buddy readout looms. Who owns this code? Why is it there? How did it get away from the team so quickly?

Many "big rewrite" projects begin with the promise of better performance. Prototypes "seem fast", but nobody's actually benchmarking them on low-end hardware. Things go fine for a while, but when sibling teams are brought in to integrate late in the process, attention to the cumulative experience may suffer. Before anyone knows it, the whole thing is as slow as molasses, but "there's no going back"... and so the lemon launches with predictably sour results.

In the midst of these crises, thoughtful organisations begin to develop a performance management discipline. This, in turn, helps to create a culture grounded in high expectations. Healthy performance cultures bake the scientific method into their processes and approaches; they understand that modern systems are incredibly complex and that nobody knows everything — and so we learn together and investigate the unknown to develop an actionable understanding.

Products that maintain a healthy performance culture elevate management of latency, variance, and other performance attributes to OKRs because they understand how those factors affect the business.

Levels of Performance Management Maturity #

Performance management isn't widely understood to be part of what it means to operate a high-functioning team. This is a communcation challenge with upper management, but also a potential differentiator or even a strategic advantage. Teams that develop these advantages progress through a hierarchy of management practice phases. In drafting this post, I was pointed to similar work developed independently by others^[3]; that experienced consultants have observed similar trends helps give me confidence in this assessment:

Level 0: Bliss #

Level 0 teams do not know they have a problem. They may be passively collecting some data (e.g., through one of the dozens of analytics tools they've inevitably integrated over the years), but nobody looks at it. It isn't anyone's job description to do so.

Folks at this level of awareness might also simply assume that "it's the web, of course it's slow" and reach for native apps as a panacea (they aren't). The site "works" on their laptops and phones. What's the problem?

Management Attributes #

Managers in Level 0 teams are unaware that performance can be a serious product problem; they instead assume the technology they acquired on the back of big promises will be fine. This blindspot usually extends up to the C-suite. They do not have latency priorities and they uncritically accept assertions that a tool or architecture is "performant" or "blazing fast". They lack the technical depth to validate assertions, and move from one framework to another without enunciating which outcomes are good and which are unacceptable. Faith-based product management, if you will.

Level 0 PMs fail to build processes or cultivate trusted advisors to assess the performance impacts of decisions. These organisations often greenlight rewrites because we can hire easily for X, and we aren't on it yet. These are vapid narratives, but Level 0 managers don't have the situational awareness, experience, or confidence to push back appropriately.

These organisations may perform incidental data collection (from business analytics tools, e.g.) but are inconsistently reviewing performance metrics or considering them when formulating KPIs and OKRs.

Level 1: Fire Fighting #

At Level 1, managers will have been made aware that the performance of the service is unacceptable.^[4]

Service quality has degraded so much that even fellow travelers in the tech privilege bubble^[4:1] have noticed. Folks with powerful laptops, new iPhones, and low-latency networks are noticing, which is a very bad sign. When an executive enquires about why something is slow, a response is required.

This is the start of a painful remediation journey that can lead to heightened performance management maturity. But first, the fire must be extinguished.

Level 1 managers will not have a strong theory about what's amiss, and an investigation will commence. This inevitably uncovers a wealth of potential metrics and data points to worry about; a few of those will be selected and tracked throughout the remediation process. But were those the right ones? Will tracking them from now on keep things from going bad? The first firefight instills gnawing uncertainty about what it even means to "be fast". On teams without good leadership or a bias towards scientific inquiry, it can be easy for Level 1 investigations to get preoccupied with one factor while ignoring others. This sort of anchoring effect can be overcome by pulling in external talent, but this is often counter-intuitive and sometimes even threatening to green teams.

Competent managers will begin to look for more general "industry standard" baseline metrics to report against their data. The industry's default metrics are moving to a better place, but Level 1 managers are unequipped to understand them deeply. Teams at Level 1 (and 2) may blindly chase metrics because they have neither a strong, shared model of their users, nor an understanding of their own systems that would allow them to focus more tightly on what matters to the eventual user experience. They aren't thinking about the marginal user yet, so even when they do make progress on directionally aligned metrics, nasty surprises can reoccur.

Low levels of performance management maturity are synonymous with low mastery of systems and an undeveloped understanding of user needs. This leaves teams unable to quickly track down culprits when good scores on select metrics fail to consistently deliver great experiences.

Management Attributes #

Level 1 teams are in transition, and managers of those teams are in the most fraught part of their journey. Some begin an unproductive blame game, accusing tech leads of incompetence, or worse. Wise PMs will perceive performance remediation work as akin to a service outage and apply the principles of observability culture, including "blameless postmortems".

It's never just one thing that's amiss on a site that prompts Level 1 awareness. Effective managers can use the collective learning process of remediation to improve a team's understanding of its systems. Discoveries will be made about the patterns and practices that lead to slowness. Sharing and celebrating these discoveries is a crucial positive attribute.

Strong Level 1 managers will begin to create dashboards and request reports about factors that have previously caused problems in the product. Level 1 teams tend not to staff or plan for continual attention to these details, and the systems often become untrustworthy.

Teams can get stuck at Level 1, treating each turn through a Development ➡️ Remediation ➡️ Celebration loop as "the last time". This is pernicious for several reasons. Upper management will celebrate the first doused fire but will begin to ask questions about the fourth and fifth blazes. Are their services just remarkably flammable? Is there something wrong with their team? Losing an organisation's confidence is a poor recipe for maximising personal or group potential.

Next, firefighting harms teams, and doubly so when management is unwilling to adopt incident response framing. Besides potential acrimony, each incident drains the team's ability to deliver solutions. Noticeably bad performance is an expression of an existing feature working below spec, and remediation is inherently in conflict with new feature development. Level 1 incidents are de facto roadmap delays.

Lastly, teams stuck in a Level 1 loop risk losing top talent. Many managers imagine this is fine because they're optimising for something else, e.g. the legibility of their stack to boot camp grads. A lack of respect for the ways that institutional knowledge accelerates development is all too common.

It's difficult for managers who do not perceive the opportunities that lie beyond firefighting to comprehend how much stress they're placing on teams through constant remediation. Fluctuating between Levels 1 and 0 ensures a team never achieves consistent velocity, and top performers hate failing to deliver.

The extent to which managers care about this — and other aspects of the commons, such as a11y and security — is a reasonable proxy for their leadership skills. Line managers can prevent regression back to Level 0 by bolstering learning and inquiry within their key personnel, including junior developers who show a flair for performance investigation.

Level 2: Global Baselines & Metrics #

Thoughtful managers become uncomfortable as repeated Level 1 incidents cut into schedules, hurt morale, and create questions about system architecture. They sense their previous beliefs about what's "reasonable" need to be re-calibrated... but against what baseline?

It's challenging for teams climbing the maturity ladder to sift through the many available browser and tool-vendor data points to understand which ones to measure and manage. Selected metrics are what influence future investments, and identifying the right ones allows teams to avoid firefighting and prevent blindspots.

Teams looking to grow past Level 1 develop (or uncover they already had) Real User Monitoring ("RUM data") infrastructure in previous cycles. They will begin to report to management against these aggregates.

Against the need for quicker feedback and a fog of metrics, managers who achieve Level 2 maturity look for objective, industry-standard reference points that correlate with business success. Thankfully, the web performance community has been busy developing increasingly representative and trustworthy measurements. Still, Level 2 teams will not yet have learned to live with the dissatisfaction that lab measurements cannot always predict a system's field behavior. Part of mastery is accepting that the system is complex and must be investigated, rather than fully modeled. Teams at Level 2 are just beginning to learn this lesson.

Strong Level 2 managers acknowledge that they don't know what they don't know. They calibrate their progress against studies published by peers and respected firms doing work in this area. These data points reflect a global baseline that may (or may not) be appropriate for the product in question, but they're significantly better than nothing.

Management Attributes #

Managers who bring teams to Level 2 spread lessons from remediation incidents, create a sense of shared ownership over performance, and try to describe performance work in terms of business value. They work with their tech leads and business partners to adopt industry-standard metrics and set expectations based on them.

Level 2 teams buy or build services that help them turn incidental data collection into continual reporting against those standard metrics. These reports tend to focus on averages and may not be sliced to focus on specific segments (e.g., mobile vs. desktop) and geographic attributes. Level 2 (and 3) teams may begin drowning in data, with too many data points being collected and sliced. Without careful shepherding to uncover the most meaningful metrics to the business, this can engender boredom and frustration, leading to reduced focus on important RUM data sources.

Strong Level 2 managers will become unsatisfied with how global rules of thumb and metrics fail to map directly into their product's experience and may begin to look for better, more situated data that describe more of the user journeys they care about. The canniest Level 2 managers worry that their teams lack confidence that their work won't regress these metrics.

Teams that achieve Level 2 competence can regress to Level 1 under product pressure (removing space to watch and manage metrics), team turnover, or assertions that "the new architecture" is somehow "too different" to measure.

Level 3: P75+, Site-specific Baselines & Metrics #

The unease of strong Level 2 management regarding metric appropriateness can lead to Level 3 awareness and exploration. At this stage, managers and TLs become convinced that the global numbers they're watching "aren't the full picture" — and they're right!

At Level 3, teams begin to document important user journeys within their products and track the influence of performance across the full conversion funnel. This leads to introducing metrics that aren't industry-standard, but are more sensitive and better represent business outcomes. The considerable cost to develop and validate this understanding seems like a drop in the bucket compared to flying blind, so Level 3 teams do it, in part, to eliminate the discomfort of being unable to confidently answer management questions.

Substantially enlightened managers who reach Level 3 will have become accustomed to percentile thinking. This often comes from their journey to understand the metrics they've adopted at Levels 1 and 2. The idea that the median isn't the most important number to track will cause a shift in the internal team dialogue. Questions like, "Was that the P50 number?" and "What does it look like at P75 and P90?" will become part of most metrics review meetings (which are now A Thing (™).

Percentiles and histograms become the only way to talk about RUM data in teams that reach Level 3. Most charts have three lines — P75, P90, and P95 — with the median, P50, thrown in as a vanity metric to help make things legible to other parts of the organisation that have yet to begin thinking in distributions.

Treating data as a distribution fundamentally enables comparison and experimentation because it creates a language for describing non-binary shifts. Moving traffic from one histogram bucket to another becomes a measure of success, and teams at Level 3 begin to understand their distributions are nonparametric, and they adopt more appropriate comparisons in response.

Management Attributes #

Level 3 managers and their teams are becoming scientists. For the first time, they will be able to communicate with confidence about the impact of performance work. They stop referring to "averages", understand that medians (P50) can tell a different story than the mean, and become hungry to explore the differences in system behavior at P50 and outlying parts of the distribution.

Significant effort is applied to the development and maintenance of custom metrics and tools. Products that do not report RUM data in more sliceable ways (e.g., by percentile, geography, device type, etc.) are discarded for those that better support an investigation.

Teams achieving this level of discipline about performance begin to eliminate variance from their lab data by running tests in "less noisy" environments than somewhere like a developer's laptop, a shared server, or a VM with underlying system variance. Low noise is important because these teams understand that as long as there's contamination in the environment, it is impossible to trust the results. Disaster is just around the corner when teams can't trust tests designed to keep the system from veering into a bad state.

Level 3 teams also begin to introduce a critical asset to their work: integration of RUM metrics reporting with their experimentation frameworks. This creates attribution for changes and allows teams to experiment with more confidence. Modern systems are incredibly complex, and integrating this experimentation into the team's workflow only intensifies as groups get ever-more sophisticated moving forward.

Teams can regress from Level 3 because the management structures that support consistent performance are nascent. Lingering questions about the quality of custom metrics can derail or stall progress, and some teams can get myopic regarding the value of RUM vs. lab data (advanced teams always collect both and try to cross-correlate, but this isn't yet clear to many folks who are new to Level 3). Viewing metrics with tunnel vision and an unwillingness to mark metrics to market are classic failure modes.

Level 4: Variance Control & Regression Prevention #

Strong Level 3 managers will realise that many performance events (both better and worse than average) occur along a user journey. This can be disorienting! Everything one thought they knew about how "it's going" is invalidated all over again. The P75 latency for interaction (in an evenly distributed population) isn't the continuous experience of a single user; it's every fourth tap!

Suddenly, the idea of managing averages looks naive. Medians have no explanatory power and don't even describe the average session! Driving down the median might help folks who experience slow interactions, but how can the team have any confidence about that without constant management of the tail latency?

This new understanding of the impact that variance has on user experiences is both revelatory and terrifying. The good news is that the tools that have been developed to this point can serve to improve even further.

Level 4 teams also begin to focus on how small, individually innocuous changes add up to a slow bleed that can degrade the experience over time. Teams that have achieved this sort of understanding are mature enough to forecast a treadmill of remediation in their future and recognise it as a failure mode. And failure modes are avoidable with management processes and tools, rather than heroism or blinding moments of insight.

Management Attributes #

Teams that achieve Level 4 maturity almost universally build performance ship gates. These are automated tests that watch the performance of PRs through a commit queue, and block changes that tank the performance of important user flows. This depends on the team having developed metrics that are known to correlate well with user and business success.

This implies all of the maturity of the previous levels because it requires a situated understanding of which user flows and scenarios are worth automating. These tests are expensive to run, so they must be chosen well. This also requires an investment in infrastructure and continuous monitoring. Making performance more observable, and creating a management infrastructure that avoids reactive remediation is the hallmark of a manager who has matured to Level 4.

Many teams on the journey from Level 3 to 4 will have built simpler versions of these sorts of gates (bundle size checks, e.g.). These systems may allow for small continuous increases in costs. Over time, though, these unsophisticated gates become a bad proxy for performance. Managers at Level 4 learn from these experiences and build or buy systems to watch trends over time. This monitoring ought to include data from both the lab and the field to guard against "metric drift". These more sophisticated monitoring systems also need to be taught to alert on cumulative, month-over-month and quarter-over-quarter changes.

Level 4 maturity teams also deputise tech leads and release managers to flag regressions along these lines, and reward them for raising slow-bleed regressions before they become crises. This responsibility shift, backed up by long-run investments and tools, is one of the first stable, team-level changes that can work against cultural regression. For the first time, the team is considering performance on longer time scales. This also begins to create organisational demand for latency budgeting and slowness to be attributed to product contributions.

Teams that achieve Level 4 maturity are cautious acquirers of technology. They manage on an intentional, self-actualised level and value an ability to see through the fog of tech fads. They do bake-offs and test systems before committing to them. They ask hard questions about how any proposed "silver bullets" will solve the problems that they have. They are charting a course based on better information because they are cognizant that it is both valuable and potentially available.

Level 4 teams begin to explicitly staff a "performance team", or a group of experts whose job it is to run investigations and drive infrastructure to better inform inquiry. This often happens out of an ad-hoc virtual team that forms in earlier stages but is now formalised and has long-term staffing.

Teams can quickly regress from Level 4 maturity through turnover. Losing product leaders that build to Level 4 maturity can set groups back multiple maturity levels in short order, and losing engineering leaders who have learned to value these properties can do the same. Teams are also capable of losing this level of discipline and maturity by hiring or promoting the wrong people. Level 4 maturity is cultural and cultures need to be defended and reinforced to maintain even the status quo.

Level 5: Strategic Performance #

Teams that fully institutionalise performance management come to understand it as a strategic asset.

These teams build management structures and technical foundations that grow their performance lead and prevent cultural regressions. This includes internal training, external advocacy and writing^[5], and the staffing of research work to explore the frontier of improved performance opportunities.

Strategic performance is a way of working that fully embeds the idea that "faster is better", but only when it serves user needs. Level 5 maturity managers and teams will gravitate to better-performing options that may require more work to operate. They have learned that fast is not free, but it has cumulative value.

These teams also internally evangelise the cause of performance. Sibling teams may not be at the same place, so they educate about the need to treat performance as a commons. Everyone benefits when the commons is healthy, and all areas of the organisation suffer when it regresses.

Level 5 teams institute "latency budgets" for fractional feature rollouts. They have structures (such as managers or engineering leadership councils) that can approve requests for non-latency-neutral changes that may have positive business value. When business leaders demand the ability to ram slow features into the product, these leaders are empowered to say no.

Lastly, Level 5 teams are focused on the complete user journey. Teams in this space can make trades intelligently, moving around code and time within a system they have mastered to ensure the best possible outcomes in essential flows.

Management Attributes #

Level 3+ team behaviours are increasingly illegible to less-advanced engineers and organisations. At Level 5, serious training and guardrails are required to integrate new talent. Most hires will not yet share the cultural norms that a strategically performant organisation uses to deliver experiences with consistent quality.^[6]

Strategy is what you do differently from the competition, and Level 5 teams understand their way of working is a larger advantage than any single optimisation. They routinely benchmark against their competition on important flows and can understand when a competitor has taken the initiative to catch up (it rarely happens through a single commit or launch). These teams can respond at a time of their choosing because their lead will have compounded. They are fully out of firefighting mode.

Level 5 teams do not emerge without business support. They earn space to adopt these approaches because the product has been successful (thanks in part to work at previous levels). Level 5 culture can only be defended from a position of strength. Managers in this space are operating for the long term, and performance is understood to be foundational to every new feature or improvement.

Teams at Level 5 degrade more slowly than at previous levels, but it does happen. Sometimes, Level 5 teams are poor communicators about their value and their values, and when sibling teams are rebuffed, political pressure can grow to undermine leaders. More commonly, enough key people leave a Level 5 team for reasons unrelated to performance management, like when the hard-won institutional understanding of what it takes to excel is lost. Sometimes, simply failing to reward continual improvement can drive folks out. Level 5 managers need to be on guard regarding their culture and their value to the organisation as much as the system's health.

Uneven Steps, Regression, & False Starts #

It's possible for strong managers and tech leads to institute Level 1 discipline by fiat. Level 2 is perhaps possible on a top-down basis in a small or experienced team. Beyond that, though, maturity is a growth process. Progression beyond global baseline metrics requires local product and market understanding. TLs and PMs need to become curious about what is and isn't instrumented, begin collecting data, then start the directed investigations necessary to uncover what the system is really doing in the wild. From there, tools and processes need to be built to recreate those tough cases on the lab bench in a repeatable way, and care must be taken to continually re-validate those key user journeys against the evolving product reality.

Advanced performance managers build groups that operate on mutual trust to explore the unknown and then explain it out to the rest of the organisation. This means that advancement through performance maturity isn't about tools.

Managers who get to Level 4 are rare, but the number who imagine they are could fill stadiums because they adopted the technologies that high-functioning leaders encourage. But without the trust, funding to enquire and explore, and an increasingly fleshed-out understanding of users at the margins, adopting a new monitoring tool is a hollow expenditure. Nothing is more depressing than managerial cosplay.

It's also common for teams to take several steps forward under duress and regress when heroics stop working, key talent burns out, and the managerial focus moves on. These aren't fatal moments, but managers need to be on the lookout to understand if they support continual improvement. Without a plan for an upward trajectory, product owners are putting teams on a loop of remediation and inevitable burnout... and that will lead to regression.

The Role of Senior Management #

Line engineers want to do a good job. Nobody goes to work to tank the product, lose revenue, or create problems for others down the line. And engineers are trained to value performance and quality. The engineering mindset is de facto optimising. What separates Level 0 firefighting teams from those that have achieved self-actualised Level 5 execution is not engineering will; it's context, space, and support.

Senior management sending mixed signals about the value of performance is the fastest way to degrade a team's ability to execute. The second-fastest is to use blame and recrimination. Slowness has causes, but the solution isn't to remove the folks that made mistakes, but rather to build structures that support iteration so they can learn. Impatience and blame are not assets or substitutes for support to put performance consistently on par with other concerns.

Teams that reach top-level performance have management support at the highest level. Those managers assume engineers want to do a good job but have the wrong incentives and constraints, and it isn't the line engineer's job to define success — it's the job of management.

Questions for Senior Managers #

Senior managers looking to help their teams climb the performance management maturity hill can begin by asking themselves a few questions:

• Do we understand how better performance would improve our business?

  • Is there a shared understanding in the leadership team that slowness costs money/conversions/engagement/customer-success?
  • Has that relationship been documented in our vertical or service?
  • Do we know what "strategic performance" can do for the business?

• What constraints have we given the team?

  • Do they have a device-class or network condition target?
  • Can engineers rely on those targets to negotiate with other parts of the organisation (e.g., sales, marketing, etc.)?

• Have we developed a management fluency wth histograms and distributions over time?

  • Do we write OKRs for performance?
  • Are they phrased in terms of marginal device and network targets, as well as distributions?

• What support do we give teams that want to improve performance?

  • Do folks believe they can appeal directly to you if they feel the system's performance will be compromised by other decisions?
  • Can folks (including PMs, designers, and SREs — not just engineers) get promoted for making the site faster?
  • Can middle managers appeal to performance as a way to push back on feature requests?
  • Are there systems in place for attributing slowness to changes over time?
  • Can teams win kudos for consistent, incremental performance improvement?
  • Can a feature be blocked because it might regress performance?
  • Can teams easily acquire or build tools to track performance?

• What support do we give mid-level managers who push back on shiny tech in favour of better performance?

  • Have we institutionalised critial questions for adopting new technologies?
  • Are aspects of the product commons (e.g., uptime, security, privacy, a11y, performance) managed in a coherent way?
  • Do managers get as much headcount and funding to make steady progress as they would from proposing rewrites?

• Have we planned to staff a performance infrastructure team?

  • It's the job of every team to monitor and respond to performance challenges, but will there be a group that can help wrangle the data to enable everyone to do that?
  • Can any group in the organisation serve as a resource for other teams that are trying to get started in their latency and variance learning journeys?

The answers to these questions help organisations calibrate how much space they have created to scientifically interrogate their systems. Computers are complex, and as every enterprise becomes a "tech company", becoming intentional about these aspects is as critical as building DevOps and Observability to avoid downtime.

It's always cheaper in the long run to build understanding than it is to fight fires, and successful management can create space to unlock their team's capacity.

"o11y, But Make it Performance" #

Mature technology organisations may already have and value a discipline to manage performance: "Site Reliability Engineering" (SRE), aka "DevOps", aka "Observability". These folks manage and operate complex systems and work to reduce failures, which looks a lot like the problems of early performance maturity teams.

These domains are linked: performance is just another aspect of system mastery, and the tools one builds to manage approaches like experimental, flagged rollouts need performance to be accounted for as a significant aspect of the success of a production spike.

Senior managers who want to build performance capacity can push on this analogy. Performance is like every other cross-cutting concern; important, otherwise un-owned, and a chance to differentiate. Managers have a critical role to forge solidarity between engineers, SREs, and other product functions to get the best out of their systems and teams.

Everyone wants to do a great job; it's the manager's role to define what that means.

If you work on a browser, you will often hear remarks like, Why don't you just put [popular framework] in the browser?

This is a good question — or at least it illuminates how browser teams think about tradeoffs. Spoiler: it's gnarly.

Before we get into it, let's make the subtext of the proposal explicit:

• Libraries provided this way will be as secure and privacy-preserving as every other browser-provided API.
• Browsers will cache tools popular among vocal, leading-edge developers.
• There's plenty of space for caching the most popular frameworks.
• Developers won't need to do work to realise a benefit.

None of this holds.

The best available proxy data also suggests that shared caches would have a minimal positive effect on performance. So, it's an idea that probably won't work the way anyone wants it to, likely can't do much good, and might do a lot of harm.^[1]

Understanding why requires more context, and this post is an attempt to capture the considerations I've been outlining on whiteboards for more than a decade.

Trust Falls #

Every technical challenge in this design space is paired with an even more daunting governance problem. Like other successful platforms, the web operates on a finely tuned understanding of risk and trust, and deprecations are particularly explosive.^[2] As we will see, pressure to remove some libraries to make space for others will be intense.

That's the first governance challenge: who will adjudicate removals? Browser vendors? And if so, under what rules? The security implications alone mean that whomever manages removals will need to be quick, fair, and accurate. Thar be dragons.

Removal also creates a performance cliff for content that assumes cache availability. Vendors have a powerful incentive to "not break the web ". Since cache sizes will be limited (more on that in a second), the predictable outcome will be slower library innovation. Privileging a set of legacy frameworks will create disincentives to adopt modern platform features and reduce the potential value of modern libraries. Introducing these incentives would be an obvious error for a team building a platform.

Entropic Forces #

An unstated, but ironclad, requirement of shared caches is that they are uniform, complete, and common.

Browsers now understand the classic shared HTTP cache behaviour as a privacy bug.

Cache timing is a fingerprinting vector that is fixed via cache partitioning by origin. In the modern cache-partitioned world, if a user visits alice.com, which fetches https://example.com/cat.gif, then visits bob.com, which displays the same image, it will be requested again. Partitioning by origin ensures that bob.com can't observe any details about that user's browsing history. De-duplication can prevent multiple copies on disk, but we're out of luck on reducing network costs.

A shared library cache that isn't a fingerprinting vector must work like one big zip file: all or nothing. If a resource is in that bundle — and enough users have the identical bundle — using a resource from it on both alice.com and bob.com won't leak information about the user's browsing history.

Suppose a user has only downloaded part of the cache. A browser couldn't use resources from it, lest missing resources (leaked through a timing side channel) uniquely identify the user. These privacy requirements put challenging constraints on the design and pratical effectiveness of a shared cache.

Variations on a Theme #

But the completeness constraint is far from our only design challenge. When somebody asks, Why don't you just put [popular framework] in the browser?, the thoughtful browser engineer's first response is often which versions?

This is not a dodge.

The question of precaching JavaScript libraries has been around since the Prototype days. It's such a perennial idea that vendors have looked into the real-world distribution of library use. One available data source is Google's hosted JS libraries service.

Last we looked, jQuery (still the most popular JS tool by a large margin) showed usage almost evenly split between five or six leading versions, with a long tail that slowly tapers. This reconfirms observations from the HTTP Archive, published by Steve Souders in 2013.

TL;DR?: Many variants of the same framework will occur in the cache's limited space. Because a few "unpopular" legacy tools are large, heavily used, and exhibit flat distribution of use among their top versions, the breadth of a shared cache will be much smaller than folks anticipate.

The web is not centrally managed, and sites depend on many different versions of many different libraries. Browsers are unable to rely on semantic versioning or trust file names because returned resources must contain the exact code that developers request. If browsers provide a similar-but-slightly-different file, things break in mysterious ways, and "don't break the web" is Job #1 for browsers.

Plausible designs must avoid keying on URLs or filenames. Another sort of opt-in is required because:

• It's impossible to capture most use of a library with a small list of URLs.
• Sites will need to continue to serve fallback copies of their dependencies.^[3]
• File names are not trustworthy indicators of file contents.

Subresource Integrity (SRI) to the rescue! SRI lets developers add a hash of the resource they're expecting as a security precaution, but we could re-use these assertions as a safe cache key. Sadly, relatively few sites deploy SRI today, with growth driven by a single source (Shopify), meaning developers aren't generally adopting it for their first-party dependencies.

It turns out this idea has been circulating for a while. Late in the drafting of this post, it was pointed out to me by Frederik Braun that Brad Hill had considered the downsides of a site-populated SRI addressable cache back in 2016. Since that time, SRI adoption has yet to reach a point where it can meaningfully impact cache hit rates. Without pervasive SRI, it's difficult to justify the potential download size hit of a shared cache, not to mention the ongoing required updates.

The size of the proposed cache matters to vendors because browser binary size increases negatively impact adoption. The bigger the download, the less likely users are to switch browser. The graphs could not be more explicit: downloads and installs fail as binaries grow.

Browser teams aggressively track and manage browser download size, and scarce engineering resources are spent to develop increasingly exotic mechanisms to distribute only the code a user needs. They even design custom compression algorithms to make incremental patch downloads smaller. That's how much wire size matters. Shared caches must make sense within this value system.

Back of the Napkin #

So what's a reasonable ballpark for a maximum shared cache? Consider:

• Browser engineers are byte misers.
• That's because any cache would need to be sized to the expected free space of the marginal (P95) user.
• Firefox for Android is between 70 and 85 MiB in compressed APK form.
• Chrome for Android varies in size between 55 and 75 MiB.
• Release-over-release growth is often less than 1 MiB.
• Cache growth will be difficult and slow.

The last point is crucial because it will cement the cache's contents for years, creating yet another governance quagmire. The space available for cache growth will be what's left for the marginal user after accounting for increases in the browser binary. Most years, that will be nothing. Browser engineers won't give JS bundles breathing room they could use to win users.

Given these constraints, it's impossible to imagine a download budget larger than 20 MiB for a cache. It would also be optional (not bundled with the browser binary) and perhaps fetched on the first run (if resources are available). Having worked on browsers for more than a decade, I think it would be shocking if a browser team agreed to more than 10 MiB for a feature like this, especially if it won't dramatically speed up the majority of existing websites. That is unlikely given the need for SRI annotations or the unbecoming practice of speeding up high-traffic sites, but not others. These considerations put tremendous downward pressure on the prospective size budget for a cache.

20 MiB over the wire expands to no more than 100 MiB of JS with aggressive Brotli compression. A more likely 5 MiB (or less) wire size budget provides something like ~25 MiB of on-disk size. Assuming no source maps, this may seem like a lot, but recall that we need to include many versions of popular libraries. A dozen versions of jQuery, jQuery UI, and Moment (all very common) burn through this space in a hurry.

One challenge for fair inclusion is the tension between site-count-weighting and traffic-weighting. Crawler-based tools (like the HTTP Archive's Almanac and BuiltWith) give a distorted picture. A script on a massive site can account for more traffic than many thousands of long-tail sites. Which way should the policy go? Should a cache favour code that occurs in many sites (the long tail), or the code that has the most potential to improve the median page load?

Thar be more dragons.

Today's Common Libraries Will Be Underrepresented #

Anyone who has worked on a modern JavaScript front end is familiar with the terrifyingly long toolchains that dominate the landscape today. Putting aside my low opinion of the high costs and slow results, these tools all share a crucial feature: code motion.

Modern toolchains ensure libraries are transformed and don't resemble the files that arrive on disk from npm. From simple concatenation to the most sophisticated tree-shaking, the days of downloading library.min.js and SFTP-ing one's dependencies to the server are gone. A side effect of this change has been a reduction in the commonality of site artifacts. Even if many sites depend on identical versions of a library, their output will mix that code with bits from other tools in ways that defeat matching hashes.

Advocates for caches suggest this is solvable, but it is not — at least in any reasonable timeframe or in a way that is compatible with other pressures outlined here. Frameworks built in an era that presumes transpilers and npm may never qualify for inclusion in a viable shared cache.

Common Sense #

Because of privacy concerns, caches will be disabled for a non-trivial number of users. Many folks won't have enough disk space to include a uniform and complete version, and the cache file will be garbage collected under disk pressure for others.

Users at the "tail" tend not to get updates to their software as often as users in the "head" and "torso" of the distribution. Multiple factors contribute, including the pain of updates on slow networks and devices, systems that are out of disk space, and harmful interactions with AV software. One upshot is that browsers will need to delete or disable shared caches for users in this state so they don't become fingerprinting vectors. A simple policy would be to remove caches from service after a particular date, creating a performance cliff that disproportionately harms those on the slowest devices.

First, Do No Harm #

Code distributed to every user is a tax on end-user resources, and because caches must be uniform, complete, and common, they will impose a regressive tax. The most enfranchised users with the most high-performance devices and networks will feel their initial (and ongoing) download impacted the least. In contrast, users on the margins will pay a relatively higher price for any expansions of the cache over time and any differential updates to it.

Induced demand is real, and it widens inequality, rather than shrinking it.

If a shared cache is to do good, it must do good for those in the tail and on the margins, not make things worse for them.

For all of the JS ecosystem's assertions that modern systems are fast enough to support the additional overhead of expensive parallel data structures, slow diffing algorithms, and unbounded abstractions, nearly all computing growth has occurred over the past decade at the low end. For most of today's users, the JS community's shared assumptions about compute and bandwidth abundance have been consistently wrong.

The dream of a global, shared, user-subsidised cache springs from the same mistaken analysis about the whys and wherefores of client-side computing. Perhaps one's data centre is post-scarcity, and maybe the client side will be too, someday. But that day is not today. And it won't be this year either.

Cache Back #

Speaking of governance, consider that a shared cache would suddenly create disincentives to adopt anything but the last generation of "winner" libraries. Instead of fostering innovation, the deck would be stacked against new and innovative tools that best use the underlying platform. This is a double disincentive. Developers using JS libraries would suffer an additional real-world cost whenever they pick a more modern option, and browser vendors would feel less pressure to integrate new features into the platform.

As a strategy to improve performance for users, significant questions remain unanswered. Meanwhile, such a cache poses a clear and present danger to the cause of platform progress. The only winners are the makers of soon-to-be obsolete legacy frameworks. No wonder they're asking for it.

Own Goals #

At this point, it seems helpful to step back and consider that the question of resource caching may have different constituencies with different needs:

• Framework Authors may be proposing caching to reduce the costs to their sites or users of their libraries.
• End Users may want better caching of resources to speed up browsing.

For self-evident security and privacy reasons, browser vendors will be the ones to define the eventual contents of a shared cache and distribute it. Therefore, it will be browser imperatives that drive its payload. This will lead many to be surprised and disappointed at the contents of a fair, shared global cache.

First, to do best by users, the cache will likely be skewed away from the tools of engaged developers building new projects on the latest framework because the latest versions of libraries will lack the deployed base to qualify for inclusion. Expect legacy versions of jQuery and Prototype to find a much established place in the cache than anything popular in "State Of" surveys.

Next, because it will be browser vendors that manage the distribution of the libraries, they are on the hook for creating and distributing derivative works. What does this mean? In short, a copyright and patent minefield. Consider the scripts most likely to qualify based on use: code for embedding media players, analytics scripts, and ad network bootstraps. These aren't the scripts that most people think of when they propose that "browsers should ship with the top 1 GiB of npm built-in", but they are the sorts of code that will have the highest cache hit ratios.

Also, they're copyrighted and unlikely to be open-source, creating legal headaches that no amount of wishing can dispell.

Browsers build platform features through web standards, not because it's hard to agree on technical details (although it is), but because vendors need the legal protections that liberally licensed standards provide.^[4] These protections are the only things standing between hordes of lawyers and the wallets of folks who build on the web platform. Even OSS isn't enough to guarantee the sorts of patent license and pooling that Standards Development Organisations (SDOs) like the W3C and IETF provide.

A reasonable response would be to have caches constrain themselves to non-copyleft OSS, rendering them less effective.

And, so we hit bedrock. If the goal isn't to make things better for users at the margins, why bother? Serving only the enfranchised isn't what the web is about. Proposals that externalise governance and administrative costs are also unattractive to browser makers. Without a credible plan for deprecation and removal, why wade into this space? It's an obvious trap.

"That's Just Standardisation With Extra Steps!" #

Another way to make libraries smaller is to upgrade the platform. New platform features usually let developers remove code, which can reduce costs. This is in the back of browser engineers minds when asked, "Why don't you just put Library X in the browser?".

A workable cache proposal features the same problems as standards development:

• Licensing limitations
• Challenging deprecations
• Opt-in to benefit
• Limits on what can be added
• Difficulties agreeing about what to include

Why build this new tool when the existing ones are likely to be as effective, if not more so?

Compression is a helpful lens for thinking about caches and platform APIs. The things that platforms integrate into the de facto computing base are nouns and verbs. As terms become common, they no longer need to be explained every time.

Requests to embed libraries into the web's shared computing base is a desire to increase their compression ratio.

Shared precaching is an inefficient way to accomplish the goal. If we can identify common patterns being downloaded frequently, we can modify the platform to include standardised versions. Either way, developers need to account for situations when the native implementations aren't available (polyfilling).

Given that a shared cache system will look nothing like the dreams of those who propose them, it's helpful to instead ask why browser vendors are moving so slowly to improve the DOM, CSS, data idioms, and many other core areas of the platform in response to the needs expressed by libraries.

Thoughtful browser engineers are right to push back on shared caches, but the slow pace of progress in the DOM (at the hands of Apple's under-funding of the Safari team) has been a collective disaster. If we can't have unicorns, we should at least be getting faster horses.

Is There a Version That Could Work? #

Perhaps, but it's unlikely to resemble anything that web developers want. First, let's re-stipulate the constraints previously outlined:

• Sites will need to opt-in.
• Caches can't be both large and fair.
• Caches will not rev or grow quickly.
• Caches will mainly comprise different versions of "unpopular", legacy libraries.

To be maximally effective, we might want a cache to trigger for sites that haven't opted-in via SRI. A bloom filter could elide SRI annotations for high-traffic files, but this presents additional governance and operational challenges.

Only resources served as public and immutable (as observed by a trusted crawler) can possibly be auto-substituted. A browser that is cavalier enough to attempt to auto-substitute under other conditions deserves all of the predictable security vulnerabilities it will create.

An auto-substitution URL list will take space, and must also be uniform, complete, and common for privacy reasons. This means that the list itself is competing for space with libraries. This creates real favouritism challenges.

A cache designed to do the most good will need mechanisms to work against induced demand. Many policies could achieve this, but the simplest might be for vendors to disable caches for developers and users on high-end machines. We might also imagine adding scaled randomness to cache hits: the faster the box, the more often a cache hit will silently fail.

Such policies won't help users stuck at the tail of the distribution, but might add a pinch of balance to a system that could further slow progress on the web platform.

A workable cache will also need a new governance body within an existing OSS project or SDO. The composition of such a working group will be fraught. Rules that ensure representation by web developers and browser vendors (not framework authors) can be postulated, but governance will remain challenging. How security researchers and users on the margins are represented are open problems.

So, could we add a cache? If all the challenges and constraints outlined above are satisfied, maybe. But it's not where I'd recommend anyone who wants to drive the web forward invest their time — particularly if they don't relish chairing a new, highly contentious working group.