The Line That Never Widened - Cover

The Line That Never Widened

Copyright© 2026 by Alexander Winston

Chapter 9: The Second Stack

May through September 1996

Curtis Boyd stood at the network map with two colored pencils, and by the time he finished, every physical link in the building was drawn twice.

Blue lines traced the TCP/IP routes carrying the bank’s new UNIX application workload. Red lines traced the IPX routes that still carried Novell login, directory, and file services. The wiring underneath hadn’t changed. The map on the wall had gotten more complicated, not less, and Curtis had drawn it that way on purpose, so that nobody looking at it could mistake the hardware consolidation for something it wasn’t.

It was the same wall he’d pinned his identity diagram to a year and a half earlier, when Janet Lewis’s transfer had first shown him how many separate systems a single employee actually touched. That diagram was long gone, replaced twice over by newer paper. The habit of drawing what the org chart preferred not to admit had outlasted every version of the map itself.


The unnamed software supplier from the spring’s certification announcement turned out to be Oracle, and its interest in UnixWare was entirely practical.

Oracle wanted to certify Oracle7 against a named UnixWare foundation release, once, and have that certification carry across both the Compaq and Dell hardware lines rather than repeating the exercise for every manufacturer that came through Novell’s door. The Compaq platform gave Oracle a credible high-end database target. Dell expanded the reach into development and departmental deployments a pure Compaq strategy would have priced out. Novell and its partner channel supplied the customers who could actually deploy the whole environment once it existed.

Oracle was careful about where its interest ended. It certified its own database against UnixWare’s operating-system interfaces and the two named hardware configurations. It had no opinion on Novell’s proprietary services, and no interest in a seat anywhere near UnixWare’s governance.

The Oracle engineer running the qualification, a methodical man named Peter Voss who had flown in from Redwood Shores expecting to spend a month arguing about compatibility, was visibly surprised at how little arguing there was to do. The same Oracle binaries installed cleanly on both machines and passed every functional test without a single conditional exception. The performance numbers told a more interesting story. The Compaq reference platform carried larger memory configurations, heavier concurrency, and recovery testing that pushed the database harder than Oracle’s own standard suite normally did. The Dell system passed cleanly too, entirely within the lower-tier capability profile it had earned in Novell’s certification lab that winter — and nothing about the shared application certification suggested the two machines were equally capable of anything beyond running the same binary correctly.

“I’ve certified against six UNIX variants this year,” Voss told Sarah, “and this is the first time I didn’t need a separate porting effort for the second vendor. Usually that’s where the month goes.”

“This is the market we wanted,” Gordon Maddix said, reading the certification summary.

“It’s proof the interfaces hold,” Eric said. “Same binary, two machines, no surprises on either side. The two-tree discipline is paying for itself.”

Oracle asked for something more than a lab result: a real customer deployment where the database, Novell’s directory, its file services, and the partner support model all had to work together at once, not as four separate certifications stacked on top of each other. Front Range was the obvious candidate. Ellen Barrett and Curtis Boyd had lived through enough of Novell’s promises by now that neither of them would mistake a successful demonstration for a production decision.


Curtis prepared the network diagram for the expanded pilot with two pencils because the traffic genuinely split into two logical worlds running over the same physical wire.

The Oracle client and its database listener spoke TCP/IP. So did the UNIX administration traffic, the backup jobs, and the new application-monitoring console Novell had bundled with the pilot. But the bank’s existing Novell clients still reached NCP file services and much of NDS over IPX, the way they always had, and every router between a branch and the data center had to carry, filter, and — when something broke — get troubleshot for both protocol families at once.

The bank’s Novell Partner supplied the CNEs handling installation and first-line support, the same support arrangement the bank had relied on all along. A Compaq system carried the primary Oracle workload; a certified Dell machine, cheaper and adequate for the purpose, ran the development and departmental testing alongside it. Dell had no role in the production database, and nobody had asked it to.

“We consolidated the hardware,” Ellen said, looking at the map. “Why do we still need two networks drawn on it?”

“Because we eliminated servers,” Curtis said, without any particular drama in his voice. “We didn’t eliminate the second routing and troubleshooting system underneath them.”

His spiral notebook came back out for the first time in months. The four identity columns from the bank’s imaging-server problem, eighteen months old now, were still there in his own handwriting. Below them he wrote two new headings: APPLICATION TRAFFIC — TCP/IP, and NOVELL SERVICES — IPX.

Sarah, standing beside Ellen, recognized the shape of it before Curtis said another word. It was the same sentence she’d used against Microsoft in a boardroom the better part of two years ago, arriving back at her from across the table. The customer was still joining together things the supplier called integrated.

“You built one operating system,” Ellen said. “You didn’t build one network.”

“No,” Sarah admitted. “Not yet.”

“The bridge moved,” Curtis said. “The toll didn’t.”


The weekend recovery test wasn’t supposed to prove anything. It was supposed to confirm the pilot was ready to expand, and instead it proved something else entirely.

A router configuration change at one branch, made in the ordinary course of the test, preserved the TCP/IP path cleanly and dropped the branch’s IPX route along with its service-advertisement traffic. The UnixWare server never noticed. Existing Oracle sessions stayed connected without interruption. A technician at the branch could ping the server and open a fresh database connection without any trouble at all.

New users could not log into Novell, could not locate the NDS services they expected to see, and could not attach to the NCP file volume they needed for their morning’s work. The path those services depended on simply wasn’t there anymore.

The first help-desk report said the UnixWare server was down. It wasn’t. Curtis demonstrated, standing at a terminal with the branch’s CNE beside him, that the server was simultaneously reachable and unavailable, and that which one was true for a given user depended entirely on which service and which transport they needed. He pinged it once for the CNE’s benefit, then opened a fresh Oracle session on the same terminal to prove the point twice, before turning to the login prompt that wouldn’t complete no matter how many times the teller retyped her password.

“Same box,” Curtis said. “Same cable. Two different answers depending on what you’re asking it.”

The CNE running the test routed the incident correctly on the first pass: the physical network was fine, the TCP/IP path was fine, the Compaq hardware was fine, Oracle was fine. The failure belonged entirely to the Novell service layer’s continuing dependence on IPX routing information that had just quietly disappeared.

Nothing was lost. No production customer was touched. That was almost beside the point. The incident reproduced, in miniature and on a Saturday, the exact ambiguity Sarah had promised the whole consolidation strategy would remove — and it had taken one ordinary router change, not a catastrophe, to surface it.

Ellen read the CNE’s incident report twice before she said anything. “If this had happened on a Tuesday, at a real branch, with real customers standing at the counter, what would they have been told?”

Nobody answered right away, which was itself the answer.

“I’m not expanding this to another branch,” Ellen said, “while we’re still running two routed environments to support one platform we’re calling integrated.”


Sarah brought Curtis’s doubled map to Eric in person rather than reducing it to a bullet point in a status deck.

Eric agreed immediately that NCP needed to run natively over TCP/IP. He was just as immediate in rejecting the idea that the fix was wrapping existing IPX packets in a TCP/IP envelope and calling the problem solved. The real list was longer than that: connection setup and recovery behavior over TCP/IP; service discovery that didn’t quietly assume IPX broadcast and advertisement traffic was still there; packet sizing, acknowledgment, timeout, and reconnect logic built for a transport with different characteristics than the one NCP had grown up on; a way for a client to choose between IPX and TCP/IP when both were available without confusing either the client or the administrator watching it; the narrow set of NDS transport changes needed for login, referral, and replica traffic; continued coexistence with the DOS, Windows 3.x, and early Windows 95 clients that had no TCP/IP-capable Novell client yet and weren’t going away on anyone’s schedule but their own; and diagnostic output specific enough that a CNE standing at a branch, not a kernel engineer in Provo, could tell which transport a failing session had actually used.

Sarah watched him write the list on the whiteboard rather than recite it from memory, which told her more about how seriously he was taking it than anything he actually said.

“NCP stays NCP,” Eric said. “File rights, authentication, connection semantics, whatever the application sees at the other end — none of that moves just because the wire underneath it changed.”

“Does the public TCP/IP stack need to change for it?” Sarah asked.

“No. The published sockets and service interfaces already give the proprietary module everything it should need.” Eric didn’t have to explain further; they’d both been in the room the winter before when that boundary got drawn. “The documented service interface we settled on during the patch dispute is the reason NCP can’t go looking for a private kernel hook this time. The public foundation supplies the networking substrate and the interface. Novell owns and supports whatever we build on top of it — same as it always has.”

He estimated the work honestly: a production-quality transition would consume engineers already committed to the first commercial Novell Enterprise UNIX release, not some pool of spare capacity sitting idle in a corner of the building.

“You’re telling me this costs the release schedule again,” Sarah said.

“I’m telling you it costs something real, the same way the last one did.” Eric didn’t dress it up. “I’d rather tell you that now than discover it in October.”


A NetWare product executive named Walter Grimes made the case against urgency in a Provo conference room a week later, and every point he raised was true.

 
There is more of this chapter...
The source of this story is StoryRoom

To read the complete story you need to be logged in:
Log In or
Register for a Free account (Why register?)

Get No-Registration Temporary Access*

* Allows you 3 stories to read in 24 hours.