Chapter 188: The Margin for Error

The time in the lower-right corner of the screen flipped to 05:12. Lin Chen had not turned on the ceiling light, leaving only a small USB lamp clipped to the edge of the monitor. Its cold white glow fell across the keyboard, picking out the thin calluses worn into his fingertips. The skeleton of the PPT was already in place, but the content was still empty. He clicked open the first page, and the cursor blinked in a steady rhythm after the title: “From Log Processing to a Data Middle Platform: A Practical Exploration of Cost and Stability.”

What he needed now was flesh. Not a pile of technical jargon, but arithmetic. What Director Li wanted was a path that could run on existing resources, not some ideal model from a lab. Lin Chen pulled up the past three months of server monitoring logs and turned CPU peaks, memory leak frequency, and disk I/O bottlenecks into line charts. The data was ugly, but it was real. He pasted the graphs into the PPT and added a note beside them: Current architecture average daily log throughput: 4.2 TB. Peak latency: 1.8 s. Mean time to recovery (MTTR): 45 minutes.

The muscle in his left calf started tightening again. He stopped typing and pushed his chair back half an inch, letting his left foot hang completely free. The pain was not sharp, but dull and heavy, like a rusted gear slowly grinding into bone. He reached into the drawer, took out a pain-relief patch, tore it open, and stuck it just below his knee. The medicinal smell mixed with the dampness that never quite left the rented room and crept into his nose. He took a deep breath and kept typing.

II. Architecture selection comparison. Option A: Deploy a Hadoop cluster directly. Pros: Mature ecosystem. Cons: Estimated hardware cost of 80,000 yuan/year, operational staffing shortfall of 2 people. Option B: Streaming processing based on the existing Kafka + Flink setup. Pros: Reuses current nodes, low transformation cost. Cons: Requires rewriting the cleansing logic; early-stage stability must be validated through canary rollout.

He chose B. There was no way back. Eighty thousand was half the department’s annual testing budget; Director Li would never approve it. He had to build a bridge over the existing mess.

At eight-thirty in the morning, his phone vibrated on the desk. The screen lit up. Old Zhao’s reply had finally come. Not a voice message this time, but a long block of text.

“Got the data. I had finance run through it once. Valid count is 11,842, same as what you reported. But more than three thousand of those are records with missing fields filled in. Do those count as valid? Also, we’ll settle the remaining payment as discussed before, but next time the format needs to be aligned in advance. The girl doing data entry on my side is clumsy—she can’t fix it.”

Lin Chen stared at the screen. Old Zhao was probing for the bottom line. Filling in empty fields was the default strategy in Lin Chen’s script, used to protect delivery speed. If Old Zhao insisted that portion did not count, a third of the final payment would be gone.

He did not reply immediately. Instead, he opened the terminal and pulled up error_log.csv and the cleansing rules document. He took three screenshots: one showing garbled text and missing fields in the raw data, one showing the script’s try-except capture logic, and one showing the completed field mapping table after fill-in. Then he typed:

“Brother Zhao, filling empty fields is a fallback strategy based on business rules. The raw data had a missing-field rate of 26%. Without completion, the downstream reports would error out directly. The fill-in logic has been written into the cleansing log and can be reviewed at any time. This batch should be settled based on the valid count of 11,842, for a remaining payment of 1,184.2 yuan. Before the next job, I’ll send you a ‘Data Access Specification Template.’ If your side fills things in according to the template, cleansing efficiency can improve by 40%, and I can lower the unit price by 5%. Does that work for you?”

He hit send. Then he leaned back in his chair and waited.

Ten minutes later, Old Zhao replied: “Fine. Send me the template. I’ll transfer the balance this afternoon.”

Lin Chen closed the chat window. He felt no extra emotion. He opened Alipay and watched the balance jump from 35.3 to 1219.5. The number had changed, but life had not. He transferred 600 to his mother with the note: Xiaoman medicine + follow-up exam. The remaining 619.5 would be kept for next quarter’s rent and more pain patches.

He knew this much: rules were not negotiated into existence. They were traded for with delivery results and viable alternatives. What Old Zhao wanted was convenience and a low price; what he wanted was cash flow and repeat business. Each side gave half a step, and only then could the transaction continue.

At eleven in the morning, the light in the rented room brightened a little. Lin Chen turned to the third section of the PPT: “Migration Cost and Risk.” This was the part most likely to trip him up. Technical reviews were not tests of coding ability—they were tests of your fallback capacity.

He remembered something Professor Zhou had said in college: “Writing code is stacking blocks. Designing architecture is building load-bearing walls. If the wall collapses, it doesn’t matter how pretty the blocks are—they’re still rubble.”

He created a new slide and drew a grayscale migration diagram. Stage 1: Dual write. Run the new link in parallel without cutting over traffic, compare only data consistency. Stage 2: Shift 10% of traffic. Observe P99 latency and error rate. Stage 3: Full switchover. Keep the old database in read-only mode for 72 hours; if anomaly thresholds are triggered, roll back automatically.

Beside it he added a bold note: Rollback is not failure; it is a safety valve. Every architecture design must include a reversible path.

He paused and rubbed the space between his brows. After sitting for five straight hours, his lower back and shoulders were as stiff as a board. He stood up and, bracing himself against the wall, slowly walked a lap around the room. The moment his left foot touched the floor, a stab of pain shot through the sole. He immediately shifted his weight back and relied on his right leg instead. His gait had already become completely distorted, but he ignored it. Pain was normal. No pain was what would be strange.

He returned to the desk and began rehearsing the defense. Facing empty air, he walked through the logic of every slide. What would Director Li ask? “How do you control costs?” “What happens if a Flink node goes down?” “If the business side won’t cooperate with the canary rollout, how do you push it through?”

He wrote the answers in a hard-cover notebook. No long essays, just keywords and numbers. Cost: Reuse existing K8s nodes, zero additional hardware. Node failure: Kafka backlog buffering + auto-restart scripts, 15-minute tolerance window. Business side: Provide a data comparison dashboard, let metrics do the talking, no arguments.

He realized he was changing—from someone who merely “took jobs” into someone who “carried responsibility.” Before, all he had to do was make the code run. Now he had to think about who would maintain it after it ran, who would take the blame if something went wrong, and where the money would come from. There was no ceremony to this shift, only dense to-do lists and deadlines that kept tightening.

At three in the afternoon, the first draft of the PPT was complete. He exported it as a PDF and sent it to his university classmate Chen Hao. Chen Hao now worked in IT operations at a state-owned enterprise in the provincial capital and had seen plenty of big-company review panels.

“Take a look at the logic for me. Just pick it apart.”

Chen Hao replied quickly: “The architecture diagram is fine. But on slide three, saying the canary period is seven days is too conservative. The business side won’t wait that long. Change it to three days, and add a line saying ‘cut over early once core metrics meet the standard.’ Also, change ‘rollback plan’ to ‘degradation strategy.’ The review committee doesn’t like hearing the word ‘failure.’ They like hearing ‘controllable.’”

Lin Chen stared at the screen, his fingers hovering over the keyboard for two seconds. Then he revised it. 7 days → 3 days. Rollback plan → degradation and circuit-breaker strategy.

He saved the file. Reformatted it. Shrunk the font by half a point and tightened the margins. He kept the whole deck under twelve slides. One page more, and it would all be nonsense.

At eight that night, his phone vibrated. An Alipay notification: Old Zhao’s remaining payment of 1,184.2 yuan had arrived. At the same time, his mother sent a voice message. He tapped it open. In the background was the sound of hospital numbers being called.

“Chen, we got the medicine. The doctor said Xiaoman’s EEG has been fluctuating a bit lately, and he should come to the provincial capital next month for another checkup. I did the math—the travel cost will be more than three hundred. On your side... if things are tight, I can borrow a little from the relatives first.”

Lin Chen listened to the message, his fingers unconsciously rubbing the edge of the desk. More than three hundred. Another piece would be cut from the cash flow he had only just pieced together. But he did not hesitate.

“Mom, no need to borrow. I’ll transfer the money. Once the follow-up date is set, send it to me. I’ll take leave and go with him.”

He hung up. Then he opened the ticketing app and checked the high-speed rail to the provincial capital next Friday. The earliest train left at 6:40 a.m. If the defense ended smoothly, he could still catch an afternoon train. But what if he got held up?

He closed the ticket page. Then he uploaded the final version of the PPT to the company’s internal approval system. The system prompt read: Submitted successfully. The review committee will complete the preliminary review by 10:00 tomorrow morning.

He shut the laptop. The room fell completely dark. His left foot began cramping again, harder this time than before. He bit down on a towel and made no sound. When the wave finally passed, he lay down slowly on the bed and stared at the water stains on the ceiling.

Tomorrow was Thursday. Once the preliminary review result came out, the tone of the final defense would be set. Old Zhao’s account was settled, the medical money had been sent, the PPT had been submitted. Every variable he could control had already been pushed onto the rails. What remained could only be waited out.

Then the phone screen lit up in the dark. A message from Director Li on the company system. The title had only four words: Preliminary Review Feedback.

Lin Chen sat up and opened it. The body was short:

“Architecture logic is clear, and the cost assessment is pragmatic. However, the ‘degradation strategy’ section lacks support from the business perspective. Before Friday’s defense, you need to add a record of alignment with the product side. Also, HR has informed us that the review committee has temporarily added an external technical advisor. Please be prepared.”

He finished reading, his finger still resting on the screen. An external advisor. Alignment with the product side.

Two more variables had just been added.

He threw back the blanket and opened the laptop again. The light from the screen once more illuminated his exhausted face. No complaints. No anxiety. Only the sound of keys striking, in the quiet rented room, again and again, tapping into the night.