Ion Stoica Kept the Lab Door Open

Before Ion Stoica learned to make a roomful of computers behave like one, he learned to draw. In Bucharest he attended art school in the mornings and regular school later in the day. He was good enough to imagine an artist's life. He was also wary of the judge. A mathematics answer could be checked; a drawing could be liked, disliked or talked around. Science offered a cleaner bargain with the world.

That childhood preference for work that can be tested survived every change of country, institution and scale. It can be found in Chord, the peer-to-peer lookup protocol he helped create. It runs through Apache Spark, which made repeated work on large datasets much faster, and Ray, which gave AI programs a flexible way to distribute jobs across machines. It is present in Arena, where models meet in blind comparisons and users vote. Stoica never stopped making things. He simply traded charcoal for code and subjective critique for benchmarks.

An education in hard edges

Stoica grew up in communist Romania, the son of two earth scientists: a geophysicist father and a geologist mother. Summers sometimes took him from Bucharest to his grandparents' village. School was organized around consequential examinations. Miss the right university place and military service waited. Stoica did serve nine months, an interval he later described without romance, though it introduced him to future classmates.

At the Polytechnic University of Bucharest, computer science won out over physics and art. He completed a master's degree in computer science and control engineering in 1989, just as Romania's political order came apart. He worked at Bucharest's Institute for Research in Informatics and began doctoral work on the side. In 1994 he moved to the United States for a Ph.D. at Carnegie Mellon.

His thesis asked how an internet router could provide fair service without remembering every flow passing through it. The proposed answer put the relevant state into packets instead. It was an early example of a Stoica habit: move complexity to the place where it can be handled, preserve scale at the center, and insist that the idea run. The work received the ACM Doctoral Dissertation Award in 2001. By then Stoica was at UC Berkeley, where he had joined the faculty the previous year.

Soon came Chord, created with Robert Morris, David Karger, Frans Kaashoek and Hari Balakrishnan. The protocol gave each item and each participating computer a position on an abstract ring, making it possible to locate data without a central directory. The diagram looked almost modest. Its importance lay in showing how order could emerge among unreliable peers. The paper earned a test-of-time award, and its ring became one of those computer-science images that outlives the machines on which it was first drawn.

Build it, then let the users complain

A paper can prove an idea. A used system can expose one. Stoica came to regard that distinction as a research advantage. Industry visitors brought the Berkeley labs problems at uncomfortable scale. Open-source releases escaped into environments the original designers had not imagined. Complaints, awkward workloads and failures returned as a better syllabus.

“You want to solve a problem ideally, which is going to be more important tomorrow than today.”Ion Stoica

The pattern was visible in Conviva, founded in 2006 with Stoica's former Carnegie Mellon adviser, Hui Zhang. Their work on delivering video over the internet met a market newly crowded with streaming. Conviva became a way to observe playback and diagnose when the experience went wrong. It also provided Stoica with a close view of slow analytics.

Meanwhile, another problem had arrived at Berkeley. Machine-learning researchers had large datasets and Hadoop, the era's standard tool, kept writing intermediate results to disk. Iterative work crawled. Stoica's doctoral student Matei Zaharia made a compact engine that kept useful data in memory. The idea became Spark. The technical insight was crisp; the route to a company was untidy.

The team discussed placing Spark with Hortonworks, a company built around Hadoop. The arrangement did not happen. In 2013 Stoica, Ali Ghodsi, Zaharia, Patrick Wendell, Reynold Xin, Andy Konwinski and Arsalan Tavakoli-Shiraji formed Databricks instead. Seven academic co-founders is a configuration designed to frighten a tidy-minded lawyer. It worked because the project already had a community, and because the founders understood different parts of the same problem.

Stoica wanted a company partly because commercial commitment would make users take the research seriously. Open source supplied reach; a company supplied continuity, support and someone to call when the cluster refused to cooperate at two in the morning. Databricks did not erase the university origin. Its early roster preserved it in plain sight, with a professor, a visiting scholar and five doctoral students sharing the founder line.

The chair he gave away

Stoica became Databricks' first chief executive. The role lasted three years. In 2016, as the company required a full-time operator, he handed the job to Ghodsi and became executive chairman. Staying in the chief executive's chair would have meant leaving Berkeley.

Silicon Valley has made a minor literary industry of the founder who abandons school. Stoica chose the less cinematic direction: the wealthy founder returning to office hours. He has explained the decision with disarming plainness. “I'm still an academic at heart,” he said. Money was not the objective; building something meaningful was. His students were not an obligation competing with the companies. They were part of the mechanism that made the companies possible.

Students, in Stoica's account, possess a useful defect. They do not reliably know what cannot be done. Experience narrows the set of plausible attempts; early confidence widens it again. His job has often been to help that confidence find a serious problem, then attach enough discipline for the answer to survive.

The record is less a family tree than a busy rail map. Zaharia went on to teach at Berkeley and serve as Databricks' chief technologist. Haoyuan Li co-founded Alluxio. Moritz helped start Anyscale. Lianmin Zheng helped build Chatbot Arena before joining UCLA. Other former students moved into faculty posts, research laboratories and new companies. Stoica's personal page lists them all with the economy of train departures: name, year, destination. The accumulation suggests that mentorship, for him, is infrastructure too.

Ray followed this route. Philipp Moritz and Robert Nishihara were trying to support AI workloads whose dynamic tasks did not fit Spark's staged rhythm. Their adviser, Michael Jordan, directed them toward Stoica's distributed-systems course. The resulting framework could schedule many kinds of work across clusters with low latency. Stoica, Moritz and Nishihara co-founded Anyscale in 2019 to support it.

The evaluator at the gate

The current list on Stoica's Berkeley page reads like a map of the modern AI stack. Spark processes data. Ray coordinates distributed AI work. SkyPilot moves workloads among clouds. vLLM serves language models efficiently. Chatbot Arena, now developed by an independent company called Arena, measures model preferences through paired human judgments. His Sky Computing Lab also asks whether cloud providers might one day interoperate more like the networks beneath them.

Yet the problem that appears to animate him now is reliability. Generative models can write software, but impressive output is not the same as a trustworthy system. Stoica asks how AI can become more predictable, accurate, verifiable and debuggable. The vocabulary has the same hard edges that drew a boy away from subjective judgment.

His honors trace the path from networks to operating systems: an early-career award from the White House, a Sloan fellowship, election as an ACM Fellow and the 2019 Mark Weiser Award. They are peer judgments about work already done. Stoica's public attention, however, tends to move quickly toward unfinished machinery. The current projects on his page are described by function, not ceremony: an inference engine, a cross-cloud launcher, an evaluation platform. Each one is a verb waiting to be tested.

His group calls one line of work AI-Driven Research for Systems. A model proposes changes to an algorithm. An evaluator runs the candidates against real workloads or simulators. The useful changes survive and generate another round. Experiments reported improvements in load balancing, inference, transaction scheduling and database operations. The model searches; the benchmark refuses to be charmed.

A deployed system is not the end of the research. It is the instrument that finds the next question.

There is an apparent provocation here. A professor whose career depends on the creativity of students is exploring machinery that can automate pieces of invention. Stoica's own framing is less theatrical. Human attention moves toward choosing consequential problems, building reliable evaluators and deciding what a result means. The tedious search can be delegated; judgment becomes more important, not less.

That brings the story neatly, perhaps suspiciously neatly, back to art school. Stoica chose a field in which answers could be tested. Then he spent a career discovering that the hard part is often designing the test. The companies, awards and valuations occupy the public ledger. The quieter continuity is a method: find a problem whose importance is rising, make an answer concrete, and leave the door open long enough for other people to show you where it fails.