Work leads, own time followsyour copy: 0 of 4 topics applied as you readlongest gap —
Porto, PortugalNine years in production (since 2017)Open to senior backend or platform roles, remote or hybrid
Arthur Costa
Senior backend & platform engineerWhat the follower saw.
Leader: my work, where production writes land first. Follower: my own time, which applies the same concerns later.
I build platforms that hold up in production: moving fleets to the cloud, turning batch into event streams, and, on my own time, execution platforms and safety tooling for coding agents.
Leader · at work writes committed in production
AC-001AC-001AC-002
replicating
Follower · own time applies them later
leader writes · follower applies
The photo is replicated the same way: writes reach the leader as green outlines and the follower applies them a moment later; then the leader dies, its unreplicated writes are marked with an X and discarded, the follower is promoted, clients retry, and both replicas converge on my face.
The replication log
one row per topic · sorted by lag, largest first
Each row is a write: a concern I first handled at work, in production, and the own-time work where I took the same concern further. Rows are sorted the way a replication dashboard sorts followers, most lag first. The pairing is mine; every fact is from my records.
Leaderat work · employers unnamed
Replication lag
Followeron my own time
Lag is the follower’s first year minus the leader’s first year, from the years on each record. Pairs are made only where both sides have a year; own-time work without one is filed under “Written straight to the follower”.
Lag and failover, drilled
two figures from two records
Two things I have built around, played as small simulations. Each figure has its text equivalent beneath it.
batch copy
CDC copy
Batch is lag
AC-002 · payment settlement, from batch to event-driven
Text equivalent. Two copies follow the same leader database. The batch copy (the saw-tooth; the shaded area under it is writes not yet replicated) falls further behind all window long and catches up only when the batch runs. The CDC copy (the thin grey line) reads the leader’s change log as it is written, so it is only ever one small step behind, at each change. Moving settlement from the first to the second is what AC-002 did; processing success rose from about 60% to 90%+. The shape is illustrative and has no scale; those two rates are the measured numbers.
client12 writes, keys k1–k12node Aleader, term 1node Bfollower
on the leader, not yet replicated
replicated
lost in failover
retry skipped: key already seen
With asynchronous replication, a leader can acknowledge writes that never reach its follower; when it fails, those writes are lost. Clients retry what they cannot confirm, and retries are safe only if each write carries a key the new leader can deduplicate. That is why my course labs reset by key (AC-007), my Firecracker driver is tested against duplicate RPCs (AC-004), and Horus treats a changed worktree as a new job, never a silent retry (AC-003).
Failover drill: retries by key
AC-007 idempotent resets by key · AC-004 duplicate RPCs in the failure suite
Written straight to the follower
no leader write behind them
In single-leader replication, followers are read-only. On my own time they are not: these writes started on the follower, one line each, the way a follower’s log reads; the last three went upstream to other projects (↗). None has a year in my records, so none shows a lag and the numbering is not a time order.
The schema both replicas run
stack
Guarantees I keep on either replica
habits
Read replicas
where others read from me
Send a write
contact
Open to senior backend or platform roles, remote or hybrid. Reach me on LinkedIn or GitHub.
leader · at work=follower · own time
In sync: both replicas hold the same state, lag zero. A write sent now lands on the leader first.
Why this design
colophon
Concept. DDIA chapter 5 in a career: work is the leader that takes production writes, my own time is the follower that replays the same concerns later, and the lag between them is measured in years from the records.
Front page. It is the replication itself: the leader’s committed writes are my results at work, arrows carry them across a lag gutter, and the follower column is my face being applied write by write, with my own-time result on top.
Information architecture. One replication log, sorted by lag like a replication dashboard: each row pairs a leader write with the follower’s applied version only where both have a year; the earlier roles sit in one outlined leader-only row, and own-time work with no leader write is filed as written straight to the follower.
Portrait and type. The ribbon portrait is the follower’s state, rows applied top-down as you read; the contact section shows both replicas in sync. Schibsted Grotesk, drawn for a newspaper, sets the headlines and text, because replication is publishing editions of one story; IBM Plex Mono is kept for positions and counters.
Colour and motion. Newsprint and ink; spruce only for writes in flight and terracotta only for writes lost in failover, always with a word or an X. Writes travel the gutter at a speed set by their lag; every figure is drawn complete at rest and replays from the start when it comes into view.
Accessibility. The ribbon’s Pause control stops the photo, the rows and both drills; reduced motion shows every end state; figures are hidden from assistive technology and each has its text beside it, and the drill writes its own log.
Complete profile
Senior backend & platform engineer · Porto, Portugal
I build platforms that hold up in production: moving fleets to the cloud, turning batch into event streams, and, on my own time, execution platforms and safety tooling for coding agents.
Open to senior backend or platform roles, remote or hybrid
Nine years in production (since 2017)
Decision records
2023–2025 · At work · Platform · Delivered
200+ instances to cloud local zones, with autoscaling
Compute cost fell by ~70% across 200+ instances.
~70% lower compute cost — 200+ instances moved from on-prem and edge sites to cloud local zones, with autoscaling
200+ instances migrated — with autoscaling, for live sporting-event peaks
Context and decision
Context. A real-time platform for live sporting events ran its workloads on on-prem and edge environments.
Decision. Lead the infrastructure modernization: migrate the workloads to cloud local zones and enable autoscaling, with multi-region load tests before peak events.
Stack. Cloud local zones · Autoscaling · Go · Python · Locust
2022–2023 · At work · Payments · Delivered
Payment settlement, from batch to event-driven
Processing success rose from 60% to 90%+.
60% → 90%+ processing success — payment settlement moved from batch to event streams with change data capture
Context and decision
Context. A legacy payment and settlement platform ran as batch processing on a monolith.
Decision. Move the core components to Java and Go, replace batch with streaming and change data capture, containerize on Kubernetes, and centralize logs and dashboards for the SLAs.
Stack. Java · Go · Streaming · CDC · Kubernetes
2026 · Own time · Platform · Working system
Horus: remote execution for builds and agent jobs
A changed worktree is a new job, never a silent retry; a worktree without a data class never leaves the machine.
Context and decision
Context. Builds and coding-agent jobs start on a workstation but need the CPU, isolation and uptime of a remote fleet, without leaking a private worktree or losing work halfway.
Decision. A local daemon with a socket API and a written durability contract; content-hash sync so a job runs on exactly the bytes submitted; recovery bound to that content; a versioned accounting ledger; gVisor sandboxes on VPS runners behind a data-class gate.
Stack. Go · Rust · gVisor
2026 · Own time · Virtualization · Spike
Firecracker microVMs as a sandbox compute driver
A failure-injection suite on a KVM host (kill mid-create, simulated reboot, duplicate RPCs, kill points 0–50 ms, a leak counter) and 23 security probes on a 22-item adversarial matrix. An unimplemented check reports SKIP, never PASS. Proposed upstream as OpenShell#4351.
Context and decision
Context. Can Firecracker microVMs, with the jailer, back an open-source agent sandbox runtime (OpenShell) through the whole lifecycle, and stay correct when things crash?
Decision. A compute driver with its own guest kernel, images from a verified catalog, reclamation of the previous run's VMs at start-up, and a guest flush before Stop.
Stack. Go · Firecracker · KVM
2026 · Own time · Platform design · Accepted
Retire my own agent platform; keep the native clients
Far less to operate; contribution rules now refuse new routers, schedulers and dashboards.
Context and decision
Context. My agent-workflow repository had grown a router, a scheduler, a coordinator hierarchy, a provider dispatcher and a dashboard in front of three coding-agent clients.
Decision. Retire it (19 October 2026). Each client runs natively; the repository keeps shared skills, a small configuration, standalone safety tools and 26 dated decision records.
Stack. Python · Bash
2026 · Own time · Platform safety · Accepted
Safety tools fail closed
Documented as a tripwire, not a sandbox, so nobody mistakes it for isolation.
Context and decision
Context. Coding agents run shell commands in real repositories from three clients with three hook systems; secrets once reached a model through process arguments.
Decision. One command-safety hook for all three clients; output redaction; long-lived credentials kept out of worker reach; a missing safety dependency refuses instead of being caught and allowed.
Stack. Python
2026 · Own time · Distributed systems · Lab
Course labs on microVMs, behind an mTLS control plane
A lab's user id stays reserved until its cleanup has succeeded; CA rotation cannot be bypassed by session resumption.
Context and decision
Context. A course platform gives every learner a disposable lab machine, created and destroyed constantly by many people at once.
Decision. A microVM lab runtime with a chaos test that kills VMs mid-session; idempotent resets by key; mutual TLS on a control-plane CA with reload without restart; pinned keyless signing that fails closed.
Stack. Go · Kotlin · protobuf
2026 · Own time · Storage & consensus · Study
A relational database, from storage to Raft
Built layer by layer as a study; each phase is a commit history.
Context and decision
Context. The fastest way to understand a distributed database is to build its layers in order and see where each guarantee comes from.
Decision. Disk manager, buffer pool, B+ tree, heap file and write-ahead log; SQL parser, planner and Volcano executor; 2PL then MVCC; Raft, Multi-Raft with range sharding, Percolator-style transactions, SSI with hybrid logical clocks, an S3 backend and the PostgreSQL wire protocol.
Stack. Rust
2026 · Own time · Sync & security · Working system
Offline-first sync with signed form-fill
The extension never fills a form the owner did not approve.
Context and decision
Context. A personal job-search tool runs on several desktops plus a browser extension; data must sync without a central source of truth.
Decision. Tombstones pruned only once every active desktop has applied them; fill tokens signed on the desktop with a P-256 key from the OS keyring; an encrypted vault export; an LLM judge measured against blind readings.
Stack. Go · Rust
2025–2026 · Own time · Developer tooling · Shipped v2.1.1
gitshift: many Git identities, one machine
Switching on a directory change went from about 3.8 s to about 0.5 s.
3.8 s → 0.5 s directory switch — gitshift identity switch on directory change, release v2.1.x
Context and decision
Context. Several GitHub and GitLab accounts on one machine means juggling keys, identities and signing by hand.
Decision. One command switches Git author, SSH key, commit signing and the GitHub CLI account, by alias or automatically by directory; releases through GoReleaser and a Homebrew tap.
Load tests to inform cost optimization and Kubernetes autoscaling; graceful shutdown and connection draining in a shared Helm chart; CI repaired with quality gates; codebase rules for agentic coding, shared with the team.
Software engineer, platform / backend
2023–2025 · Real-time platform for live sporting events
Led the move of on-prem and edge workloads to cloud local zones with autoscaling (AC-001); real-time, event-driven services on serverless and containers; Go and Python automation; Locust multi-region load tests for peak events; CI/CD across GoCD, GitHub Actions, Jenkins, Buildkite, Ansible and Packer; metrics, logs, traces and alerts.
Software engineer, payments / settlement
2022–2023 · Payment and settlement platform (consulting)
Led the modernization of a legacy payment and settlement platform to event-driven streaming and CDC (AC-002).
Software engineer, post-purchase
2021–2022 · E-commerce marketplace, several countries
Backend services in Go, Java, Node.js and Groovy for post-purchase and payment workflows; faster claim and dispute resolution.
Software engineer
2019–2021 · Core banking, Brazil
Monoliths and mainframe-based systems modernized into Java and Spring microservices; migration of microservices to the public cloud under regulatory SLAs.
Associate software engineer, CRM & integrations
2018–2019 · Fintech CRM
Real-time CRM features integrated with banking services such as loans, OTP and account management.
Associate software engineer
2017–2018 · Software house
CRM and warehouse management systems in PHP and JavaScript.
How I work
Fail closed: a missing safety dependency refuses; it is never caught and allowed.
Crash and recovery suites before claims: kill mid-create, simulated reboots, leak counters.
Evidence tied to a commit; a check that is not implemented says SKIP, never PASS.
Decision records for every significant choice, including retiring my own platform.
Load tests before peak events.
Also built
Drover
Watch and approve agent sessions from any device: a Rust core over UniFFI, pairing scopes bound to the consumed token.
Rust
JVM performance labs
Eleven on-heap versus off-heap lock-free scenarios with JMH, and 13 serialization frameworks behind one endpoint. Numbers not quoted until re-run with raw output.
Java · JMH
graphql-federation
Federation 2 across Java, Go and TypeScript subgraphs, gRPC between services, SLO burn-rate dashboards.
Java · Go · TypeScript · gRPC
Ponte
A management system for tutoring schools: live lessons with recording and consent, TOTP and lockout, a finance ledger ported to Go and checked by replaying a shared corpus.
TypeScript · Go · PostgreSQL
deep-engineering-knowledge-base
An LLM pipeline that writes and checks engineering lessons, with an independent critic that excludes the author.
Python
techish-visual
An engine for animated technical diagrams: a ThorVG WASM renderer and performance budgets in CI.
TypeScript
keyframe
A Linux studio for recording and editing technical videos, with cuts snapped to silence.
Rust
java-workflows
Reusable GitHub Actions for Java teams: a Java 11 to 25 matrix, security scanning and SBOMs.
GitHub Actions
Upstream
OpenShell#4351
Firecracker compute driver proposal, from AC-004 (open source). · open
OpenShell#4591
A gateway reporting "database is locked" under two concurrent clients. · open
Orca, 10 pull requests
SSH OAuth forwarding, pairing-offer rotation without restart, security events, isolation from external agent config. · 9 open, 1 closed