Backend on Osmar Petry

Backend Architectures — history, case studies & dogfooding

Mon, 01 Sep 2025 00:00:00 +0000

Summary

Know the history (monolith → services → serverless) to choose the right shape for now.
Use dogfooding to turn internal capabilities into platforms and eventually products.
Pair tech + finance on every decision so architecture choices have a clear P&L story.

Case studies (what they did → why it mattered)

Amazon: built infra & tooling for itself first (dogfood) → later externalized as AWS (new revenue, 2× monetization of the same effort).
Netflix: hybrid of microservices + serverless; serverless for bursty/edge workloads, services for core domains → speed + cost control.
Google: evolved from large coordinated monoliths to microservices as scale/teams grew → independent deploys and targeted reliability.
Uber: reused profile, payments, geo, notifications to launch Eats quickly → time-to-market by reusing platform primitives.

Decision heuristics (when to pick what)

Monolith first when the team is small, domain still moving, and you need velocity.
Split by domain (bounded contexts) when teams block each other and deploys become risky.
Serverless for event-driven, spiky, or edge tasks where operational burden must be minimal.
Always design for observability (logs/metrics/traces) and clear failure domains.

Dogfooding → productization loop

Build the capability you need (internal API/service).
Harden with SLOs, docs, dashboards.
Reuse across 2–3 internal products (prove generality).
Abstract contracts (versioning, quotas, cost model).
Externalize (SDKs, billing, support) when economics are clear.

Two-sided payoff: internal reuse lowers cost/time, externalization creates revenue.

Modularization Criteria — Parnas 1972

Tue, 15 Apr 2025 00:00:00 +0000

Abstract

David L. Parnas (1972) proposed information hiding as the primary criterion for modular design, using the KWIC (Key Word In Context) system to compare traditional flow-based modularization vs encapsulating design decisions. His approach yields systems that are easier to modify, maintain, and evolve.

Key points

Traditional modularization (aligned with processing steps) leads to high coupling and brittle systems.
Information hiding: modules encapsulate design decisions (data structures, algorithms, storage formats).
KWIC case study: conventional design (5 modules aligned with processing sequence) vs information-hiding design (4 modules for storage, circular shift generation, sorting, output).
Benefits: localized changes, clear responsibilities, replaceable internals without breaking clients.
Trade-off: potential performance overhead due to inter-module communication; can be mitigated via inlining and efficient call mechanisms.

Connections

Precursor to object-oriented encapsulation and modern modular design.
Supports microservices and DDD bounded context thinking.
Relates to clean architecture and low coupling principles.
Anticipates interface contracts in modern API design.

Questions

How to apply information hiding in fast-paced agile projects?
Which metrics can quantify effective information hiding?
How to balance encapsulation with performance in real-time systems?
Applicability to distributed architectures?

Personal reflections

Encapsulating design decisions prevents leakage of implementation details across layers.
KWIC example illustrates how storage strategies can remain swappable with proper abstraction.
Critique of flow-based modularization still relevant for ETL/data pipeline architectures.
Stable interfaces support DevOps pipelines and isolate testing concerns.

References

Parnas, D. L. (1972). On the Criteria To Be Used in Decomposing Systems into Modules. Communications of the ACM, 15(12), 1053-1058.
Modern commentary: https://www.refact0r.dev/blog/module-criteria

[[conways-law-birth-1968|Conway’s Law — origins, literature, and team design]]
[[DbC|Design by Contract]]
[[no-silver-bullet-1986|No Silver Bullet — Essence & Accidents]]
[[software-aging-1994|Software Aging]]

Master’s Article Summaries

Wed, 04 Sep 2024 00:00:00 +0000

TL;DR

Replay attacks remain the easiest ASV spoofing vector; DL-RAD, autoencoders + Siamese networks, and CQCC features significantly improve detection.
ASVspoof 2021 pushes detection into real-world, noisy settings requiring domain generalisation.
Remote sensing (NEON/NIST) mirrors the need for multi-source data fusion—hyperspectral, LiDAR, RGB—for ecological insights.

SPIED articles

Ren et al. — Replay attack detection via loudspeaker distortion (DL-RAD).
ASVspoof 2021 — Spoofed/deepfake speech detection in the wild.
NIST DSE — Plant identification with airborne remote sensing.
Adiban et al. — Autoencoder + Siamese countermeasures on ASVspoof 2019.

Context

Consolidated notes from 4 September 2024 research sprint.
Focus: voice authentication security (spoofing/deepfake) and ecological remote sensing.
Supporting docs: ZIP archive with slides/text; online share for extended summaries.

1. Replay Attack Detection Based on Distortion by Loudspeaker

Source: Ren et al., Multimedia Tools and Applications, 2019. DOI: 10.1007/s11042-018-6834-3.
TL;DR: DL-RAD detects replay attacks by analysing loudspeaker-induced distortions (low-frequency attenuation, harmonic energy).
Highlights: Harmonic Energy Ratio, Low Spectral Variance. Achieves >98% detection accuracy.
Application: voice authentication systems (mobile, banking). Focus on dependable feature extraction.
Reflection: Consider how speaker hardware signatures can serve as anti-spoof signals.

2. ASVspoof 2021: Deepfake Speech Detection in the Wild

Source: ASVspoof 2021 challenge; TASLP 2023 paper (DOI: 10.1109/TASLP.2023.3285283).
TL;DR: Evaluates spoofed/deepfake detection in noisy, uncontrolled environments; introduces large-scale dataset.
Highlights: Variance across capture devices, environmental noise; combination of spectrogram analysis and deep models.
Application: deploy robust detectors for real-world ASV systems, banking, call centers.
Reflection: emphasises the need for adaptive models and domain generalization.

3. NIST DSE Plant Identification with Remote Sensing

Source: NIST publication on airborne remote sensing data challenge.
TL;DR: Integrates hyperspectral, LiDAR, RGB data to segment tree crowns, align field data, classify species.
Highlights: data fusion, scaling ecological monitoring, addressing heterogeneous resolutions.
Application: environmental monitoring, conservation, precision agriculture.
Reflection: parallels with multi-modal data integration in other domains (e.g., security sensors).

4. Replay Spoofing Countermeasure Using Autoencoder & Siamese Networks (ASVspoof 2019)

Source: Adiban et al., Computer Speech & Language, 2020. DOI: 10.1016/j.csl.2020.101105.
TL;DR: Combines autoencoders (denoising) with Siamese networks (similarity) to detect replay attacks.
Highlights: CQCC features, improved EER by 10.73%, t-DCF drop of 0.2344.
Application: mobile authentication, payment systems, secure access.
Reflection: underscores the power of hybrid feature + metric-learning approaches.

Supporting Docs

Comparative notes across TXT/PPTX/Google Docs for detailed methodology.
Extended definitions (CQCC, EER, t-DCF) stored in local dictionary.
Presentations (March 25) outline challenge evolutions and future work.

Next steps

Investigate combined defenses against multi-modal spoofing (synthetic + replay).
Explore edge deployment viability for real-time detection.
Compare ecological data pipelines with security workflows for cross-domain insights.