Steve McCanne and the trouble with invisible things

A synthesizer is an agreeable place to begin a career in computer networking. It makes a noise when you do something right, and often when you do something wrong. For Steve McCanne, an interest in making music helped open a door into digital signal processing. The rock-star ambitions belonged to high school. The technical curiosity found a considerably longer engagement.

His subsequent career accumulated the names that appear in technology biographies: FastForward Networks, Inktomi, Riverbed, Jut, Brim Data. Yet the company list can obscure the interesting continuity. Again and again, McCanne has worked on the awkward distance between a computer system and the person trying to understand it. Sometimes the distance is measured in network traffic. Sometimes it is buried in the shape of the data.

Following that thread produces a more revealing portrait than following the financing rounds alone. It also explains why a founder with several companies behind him still uses a job title with an unusually specific first word: coding.

The keyboard before the keyboard

McCanne earned his Berkeley undergraduate degree in 1990 and his doctorate in 1996. His early professional world included Lawrence Berkeley National Laboratory, where networking research meant working on systems that other researchers needed to use. The internet was something to investigate, instrument, and improve. Looking inside it required tools as well as theories.

The laboratory’s network research community included Van Jacobson, Craig Leres, Vern Paxson, and Sally Floyd. McCanne’s work belongs in that collaborative setting. The tools associated with his name grew out of people solving practical problems together, rather than a solitary revelation followed by a company logo.

A better way to watch the wire

Consider a basic difficulty in examining a network. An application wants to study traffic, but moving every available packet into the application creates work of its own. The act of observing can become expensive. A useful monitoring system has to be selective about what it passes along.

In their 1993 BSD Packet Filter paper, McCanne and Jacobson described a design that filtered unwanted packets early and used a register-based evaluator. They also paid attention to buffering. These are the sort of details that disappear from a product announcement and matter enormously when the product must actually run.

Packet capture sounds narrow until you need to find out what a network is doing. Then the ability to collect the right evidence becomes the beginning of every useful conversation. McCanne’s work on pcap and his co-authorship of tcpdump sit in that practical tradition: build a way for other people to investigate their own problems.

Video presented a related challenge at a different scale. McCanne worked on MBone tools and multicast communication, where a transmission could reach multiple participants. His doctoral work examined scalable compression and transmission of internet multicast video. He joined the Berkeley faculty in 1996; ACM lists his dissertation award under 1997.

The research question had consequences beyond the laboratory. If people in different places were going to collaborate through audio, video, and shared tools, the machinery had to cope with the network they actually had. McCanne’s early work combined signal processing with the less glamorous business of making communication function.

Coffee, code, and a public company

In 1998, McCanne co-founded FastForward Networks, turning toward internet broadcasting. Inktomi acquired the company in 2000 in a deal reported at $1.3 billion, and he became Inktomi’s chief technology officer. It was a large corporate transaction attached to a very particular engineering subject: distributing live media over the internet.

Riverbed began in 2002 with Jerry Kennelly. The origin included a Saturday meeting at Starbucks, a setting sufficiently ordinary to resist the usual startup mythology. Kennelly later recalled a straightforward division of labor: he prepared the business plan while McCanne developed the software. One could explain the enterprise to investors; the other could make the underlying idea work.

They were tackling the frustrations of wide-area networks. A company could connect distant offices and still leave people waiting for applications and files. Geography, repeated transfers, and protocol behavior all had a vote. Buying a connection did not settle every argument about performance.

Riverbed’s Steelhead appliances addressed that problem. Commercial shipments began in May 2004, and the company completed its initial public offering in September 2006. The progression from founding to shipping to a public market makes a compact timeline. Each date also represents a different test: whether an idea can be built, whether customers want it, and whether an organization can deliver it repeatedly.

McCanne remained Riverbed’s CTO until 2013. Along the way, Berkeley recognized him with its 2009 CS Distinguished Alumni Award. The academic and commercial chapters had become thoroughly intertwined. Research could suggest a design; a customer’s network supplied an exacting examination of it.

The monster assembled from useful tools

At Jut, founded in 2013, McCanne turned his attention toward operational data. A team might have logs in one place, metrics in another, and events somewhere else. Each tool could be useful. Getting them to answer a question together could become a small construction project.

Customers had a name for the assembled creature: FrankenOps. The joke works because the parts can be perfectly respectable. The trouble begins at the seams. A visualization in one product and a relevant record in another do not automatically become a coherent explanation.

Jut announced a $20 million Series B in November 2013, led by Accel with Lightspeed and Wing participating. Its operations data hub entered open beta in 2015. Juttle, its dataflow language, was intended to connect questions about live streams with questions about historical information. McCanne’s ambition extended to applications sharing the same underlying data platform.

There was a business argument underneath the tooling. If software increasingly ran a company’s operations, understanding that software became a concern for more people than the developers. Jut aimed to give developers and operations teams a foundation from which other users could explore what was happening. The intended audience could expand without requiring every person to learn the implementation underneath.

That was Jut’s stated vision at the time. It is also a useful bridge between McCanne’s networking work and his later interest in data models. The person asking the question should be able to change the question. How much rebuilding should that require?

A flamingo with an identity problem

McCanne’s technical writing offers a view of how he explains these difficulties. A 2022 tutorial on Go serialization introduces a rose and a flamingo. Both can have a color. Turn their values into JSON, however, and information about which kind of thing they were can disappear. Preserving a property has failed to preserve the whole meaning.

It is a pleasant example because the problem arrives before the terminology. A reader can see the missing information and then follow the proposed solution. McCanne uses ZSON to show how type information can travel with the data. The little bird is doing substantial educational work.

“Is this the best we’ve got?”

Steve McCanne, on the complications of Go deserialization, 2022

His writing about zq similarly begins with the difficulty of using an existing tool, jq, while acknowledging its usefulness. The concern is how easily a person can reason about the computational model once a copied example stops being enough. A language can express a great deal and still make its user work hard to understand what happens next.

In 2023, McCanne co-authored the CIDR paper Zed: Leveraging Data Types to Process Eclectic Data with Amy Ousterhout, Henri Dubois-Ferrière, Silvery Fu, Sylvia Ratnasamy, and Noah Treuhaft. The paper argued for making types central to handling data with varied, evolving schemas. Once again, the design question sat underneath the visible application.

His public career also keeps returning to collaboration across generations of tools. The colleagues around the early network work and the co-authors of the later data research make those connections concrete. A useful design has to be explainable to another person, whether that person is implementing a filter, reading a paper, or deciding what a query means. The shared work gives this career its texture: individual technical contributions, carried into systems that other people can test and extend.

Still close to the code

Brim’s history provides another connection. Phil Rzewski, now its technical director, described seeing an early prototype after years of working with McCanne on and off. He joined in late 2018. The invitation was to look at a demonstration: a working idea that could make the opportunity tangible.

Brim Data now lists McCanne as its Coding CEO. His public GitHub profile names SuperDB alongside his earlier packet tools. SuperDB’s stated direction is to bring relational tables and varied JSON into a shared data model, with self-describing values and strong typing. The aspiration remains technical and practical: reduce the awkwardness of working across different forms of information.

His GitHub biography ends with another occupation: rancher. LinkedIn places him in Fall River Mills, California. Those details give a little breathing room to a career otherwise crowded with protocols, languages, and company names. An executive biography generally finds space for another board seat. McCanne’s makes room for ranching.

The recurring interest is visible without pretending every project was part of a plan written decades ago. A synthesizer led toward signals. Network research produced tools for inspecting traffic. Companies took on broadcasting, application performance, and analytics. The work kept finding new forms of friction. McCanne kept getting close enough to the machinery to try another design.