Software performance optimization
Find what slows the work. Agree on what an improvement must prove.
Slow screens, delayed exports and growing job queues interrupt real operations. Adding servers may help, but it can also leave the underlying delay in place.
Bring Better Software one slow workflow. We can discuss a scoped engineering investigation: define the workload, establish a baseline, identify the evidence needed and agree on the next change. Implementation scope follows what the investigation supports.
Thirty minutes with Jai, Better’s founder. Bring the operation, its business deadline and the measurements you already have.
01
When this work makes sense
Customers wait for an operation that used to finish quickly. A background job misses an operating deadline. Performance changes as a customer’s data grows. Your team has plausible fixes but cannot yet show which constraint matters.
The starting point is one operation with a clear owner. A general request to make the entire application faster needs narrowing before it can become a useful engagement.
If you need incident command for an active outage, round-the-clock response or specialist database internals work, say so at the start. We would confirm the required people and scope before committing. This page does not promise emergency coverage or specialist availability.
02
Measure the finished operation
An API can acknowledge a request quickly while the resulting export sits in a queue. A query can finish sooner while the customer still waits for the file. Define where the operation starts, where it ends and what makes its output correct.
Google’s SRE monitoring guidance treats latency, traffic, errors and saturation as separate signals. That is a useful starting discipline: faster failed requests cannot stand in for a better customer experience.
The investigation should connect the operating delay to supporting measurements. Brendan Gregg’s USE method checks utilization, saturation and errors across resources. It is one diagnostic method with limits, not a guarantee that a resource chart will reveal every application problem.
Customer elapsed time spans the complete operation
- 01Submit
API acknowledgement can end here
- 02Queue
- 03Process
- 04Validate output
- 05Available
03
A scope you can review before changes begin
Stage
Define
Work to agree
One workflow, its deadline, expected output and representative demand
Evidence for the decision
A written workload and acceptance boundary
Stage
Measure
Work to agree
Application path, queue time, relevant dependencies and resource limits
Evidence for the decision
Baseline results, environment and measurement gaps
Stage
Investigate
Work to agree
Ranked explanations and the smallest useful experiment
Evidence for the decision
Traces, profiles or query evidence supporting each hypothesis
Stage
Change
Work to agree
An agreed implementation, test conditions and rollback procedure
Evidence for the decision
Reviewed change record and comparable results
Stage
Observe
Work to agree
Production checks and ownership after release
Evidence for the decision
Remaining limits and a handover record
A diagnosis and an implementation are different commitments. Agree on each deliverable and its budget. If the available evidence cannot reproduce the problem, the useful result may be an instrumentation plan rather than an unearned root-cause claim.
04
Example: an export that looks fast but fails the deadline
This is an illustrative decision model, not a Better customer result.
Suppose 100 exports are submitted during an agreed test. Ninety become available in 20 seconds; ten remain unfinished at the observation cutoff. A report containing only completed exports makes every recorded completion look fast. It still cannot show that 95% of submissions met a 60-second deadline. Only 90% are known to have done so.
Use the acceptance sheet below to make that distinction explicit before work begins. The business chooses its actual targets; the example numbers are not recommendations or promises.
Acceptance field
Start and finish
Illustrative entry
Export submitted to correct file available for download
Acceptance field
Data
Illustrative entry
Fixed authorized test dataset; record size and tenant distribution
Acceptance field
Demand
Illustrative entry
Two export submissions per minute plus an agreed interactive traffic mix
Acceptance field
Deadline
Illustrative entry
At least 95% of submissions finish within 60 seconds
Acceptance field
Outcome accounting
Illustrative entry
Count failures, timeouts and unfinished work separately; none disappear from the submitted population
Acceptance field
Correctness
Illustrative entry
Completed files pass the agreed record and field comparison
Acceptance field
Comparison boundary
Illustrative entry
Record software version, worker/database allocation, cache state, duration and observation cutoff
Acceptance field
Release decision
Illustrative entry
Named approver, remaining gaps and rollback owner
The acceptance sheet is a starting specification. It is not a benchmark or a test harness. A short test cannot establish long-term reliability or behavior at untested demand.
05
Choose the change with the cost visible
A code or query change may reduce unnecessary work. More capacity may be the fastest practical option. Moving work into a queue may suit a background operation, but the queue still needs a completion deadline and failure handling. A product limit may be preferable to paying for a workload the business rarely needs.
Compare these choices against the same operating target. Include infrastructure cost, engineering effort, correctness and maintenance. Avoid accepting a response-time gain that depends on serving stale data or removing required validation.
Representative testing matters. Grafana’s load-model documentation explains that in a closed model, slower iterations can reduce the incoming work rate. Select a test model that matches the real demand and record whether the load generator delivered it. A smaller workload is not evidence of better capacity.
06
What we need from your team
Bring the slow workflow, affected users, business impact, stack and current measurements. Useful inputs include recent changes, data growth, logs and a safe way to reproduce the behavior. Share only material your team is authorized to provide.
Before investigation, agree on access, production restrictions and who may approve changes. Database analysis must fit the engine and environment. PostgreSQL’s EXPLAIN guidance notes that EXPLAIN ANALYZE executes the query, including side effects. Diagnostic tooling needs an agreed safe environment and an engineer qualified for the actual stack.
Better’s broader engineering work is described on our about page. We do not present that as proof of a specialist profiling credential or a promised speedup. Team allocation, responsibilities and budget are agreed before work begins; see how our engineering engagements are priced.
07
Discuss the workflow and the decision it blocks
Bring one slow operation, its deadline and what your team already knows. We will discuss whether a scoped engineering investigation is the right next step and what must be confirmed before an engagement.