DevOps Services in Barcelona, Spain

Tourism / Travel Tech / DevOps / Barcelona

A Barcelona-based travel technology company (EUR 180M revenue, operating a vacation experience marketplace connecting 2.4M annual travellers with 45,000 experience providers across 120 destinations in Europe and Latin America, with 420 employees and a platform processing 85,000 daily bookings during peak season) needed DevOps transformation to scale beyond infrastructure that buckled every summer. The company had grown 55 percent YoY for 3 years, but each summer brought infrastructure crises that damaged customer experience and provider relationships. Core challenges: (1) Summer scaling catastrophe — the platform experiencing radical seasonal variation: January (low season) processing 12,000 daily bookings versus August (peak) processing 85,000 daily bookings — a 7x demand swing. The infrastructure was provisioned for "average" demand (approximately 40,000 daily bookings) with manual scaling adjustments as summer approached. In summer 2024: peak traffic exceeded provisioned capacity on 18 separate days, causing: search response times degrading from 800ms to 12 seconds (users abandoning after 3 seconds), booking failures increasing from 0.8 percent to 6.2 percent during peak hours, and provider availability feeds lagging 15-45 minutes behind real-time (causing overbookings and customer complaints). Estimated revenue impact: EUR 8.4M in lost bookings during the 18 degraded days. The engineering team spending June through September in perpetual crisis mode — fixing infrastructure fires rather than building product features. (2) Provider integration fragility — 45,000 experience providers connected through a mix of: direct API integration (8,000 large providers), channel manager APIs (15,000 medium providers through 12 different channel managers), and manual inventory management (22,000 small providers uploading through a web portal). Each integration having different availability feed formats, booking confirmation protocols, and cancellation policies. Deployment of integration updates was high-risk — a bug in the channel manager integration could cause overbookings across thousands of providers simultaneously. The integration layer deployed monthly with a 4-hour maintenance window — during which no bookings could be processed. Each monthly deployment carrying significant risk (12 percent failure rate requiring rollback). (3) Multi-currency and multi-language complexity — the platform operating across 22 currencies and 12 languages, with pricing rules varying by market (Spanish VAT different from French TVA different from Mexican IVA), cancellation policies varying by jurisdiction (EU consumer rights applying in European markets, different rules in Latin American markets), and content (experience descriptions, reviews, provider communications) requiring management across all languages. Deployment of pricing or policy changes was terrifying — a mistake in currency conversion or tax calculation could result in thousands of incorrectly priced bookings in minutes. The team deploying pricing changes during off-hours with a 3-person verification team watching real-time transactions — a manual process that didn't scale. (4) Mobile performance — 72 percent of bookings coming from mobile devices (with 45 percent from the native app and 27 percent from mobile web). Mobile performance was critical: app store ratings directly correlated with booking conversion — a 0.1-point rating decrease corresponding to approximately 2 percent conversion decline. The mobile app deployed every 2 weeks — iOS App Store review adding 2-5 days of unpredictable delay. Android updates via Google Play deploying faster but still requiring manual staged rollout. Backend API changes requiring coordination with mobile releases — a backend change that broke the mobile API causing app crashes and immediate rating damage. (5) Development team scaling pain — the engineering team growing from 40 to 120 engineers in 2 years, but development velocity had not increased proportionally. New engineers waiting 3-5 days for development environment setup. Code merge conflicts increasing as team size grew — the monolithic codebase meaning everyone working in the same repository. Deployment queue: with monthly integration layer deployments and bi-weekly mobile deployments, feature delivery was bottlenecked — features completed in March sometimes not reaching production until June.

We delivered DevOps transformation over 12 weeks — solving the seasonal scaling crisis and enabling the engineering team to deliver features at the pace the business required. (1) Seasonal auto-scaling: infrastructure that breathes with demand. Kubernetes platform: all platform components migrated to EKS with auto-scaling policies tuned for travel seasonality. Three scaling dimensions: HPA (Horizontal Pod Autoscaler) for request-based scaling (API servers scaling with concurrent users), KEDA for event-driven scaling (booking processors scaling with queue depth), and Cluster Autoscaler for node-level capacity. Predictive scaling: machine learning model trained on 3 years of traffic patterns — predicting demand 72 hours ahead based on: day of week, season, holidays across all markets (Spanish, French, British, German, Mexican holidays all creating different demand patterns), weather forecasts for key destinations, and marketing campaign schedules. Infrastructure pre-scaled before demand arrived rather than reacting to traffic spikes. Cost-optimised capacity: mixed instance strategy — Reserved Instances for baseline capacity (approximately 40 percent of peak), Spot Instances for elastic capacity with automatic fallback to On-Demand. Infrastructure cost per booking reduced 52 percent through right-sizing and eliminating the "provision for peak year-round" approach. Summer 2025 infrastructure cost: EUR 380K (versus EUR 680K in summer 2024) while handling 35 percent more bookings. (2) Integration layer modernisation: safe, independent provider deployments. Microservice decomposition: the monolithic integration layer decomposed into services per integration type — direct API connector service, channel manager service (one adapter per channel manager), and web portal service. Each independently deployable, testable, and rollback-able. Provider isolation: each provider integration running in an isolated execution context — a bug in one channel manager adapter affecting only providers connected through that channel manager, not the entire platform. Circuit breaker patterns preventing provider API failures from cascading to the booking platform. Zero-downtime deployment: monthly maintenance windows eliminated — integration updates deployed through canary releases during business hours. New integration versions tested with shadow traffic (real provider requests processed by both old and new versions, with only old version responses used) before progressive cutover. Automated rollback within 30 seconds if booking success rate drops below threshold. Webhook reliability: outbound notifications to providers (booking confirmations, cancellations, modifications) delivered through a reliable event queue with guaranteed delivery and retry logic — eliminating the "lost booking" issues that plagued provider relationships. (3) Multi-currency and multi-language safety: configuration-as-code with validation. Configuration management: pricing rules, tax rates, cancellation policies, and currency conversion rates managed as versioned code — changes going through the same CI/CD pipeline as application code, with automated testing validating: tax calculation correctness across all 22 currencies, cancellation refund calculations per jurisdiction, and currency conversion accuracy against reference rates. Pricing sandbox: pricing changes deployable to a shadow environment that processes real bookings through the new pricing logic without affecting actual charges — engineering team verifying that new pricing produces expected results across a statistically significant sample of real transactions before production activation. Feature flags for market-specific changes: regulatory or pricing changes in one market deployable without affecting other markets — French TVA rate change rollout tested and deployed independently of Spanish or German pricing. Content pipeline: multilingual content (descriptions, reviews, provider communications) managed through a deployment pipeline with automated quality checks — missing translations flagged, formatting issues detected, and cultural appropriateness verified through automated rules. (4) Mobile deployment acceleration: fast, safe app releases. Backend API versioning: API versioned to support multiple mobile app versions simultaneously — new backend features deployable without breaking older app versions. API deprecation lifecycle: new version deployed, old version maintained for 90 days, gradual migration as users update apps. App deployment pipeline: automated build, test, and submission for iOS and Android — pull request merged to release branch triggering: platform-specific builds, automated UI testing (visual regression, functional testing on device farms), App Store Connect submission (with automated metadata and screenshot generation), and Google Play staged rollout (1 percent, 10 percent, 50 percent, 100 percent with automated monitoring). Release frequency: from bi-weekly to weekly iOS releases and twice-weekly Android releases (Android having faster store approval). Emergency fix pipeline: critical bug fixes deployable to both platforms within 24 hours (including App Store review through expedited process). App performance monitoring: real-time mobile APM integrated into the observability platform — crash rates, cold start times, API response times, and user flow completion rates monitored continuously. App store rating tracked as a DevOps metric — degradation triggering investigation. (5) Developer experience: fast onboarding and independent delivery. Environment automation: new engineer development environment provisioned in 45 minutes (from 3-5 days) — automated script setting up: local Kubernetes cluster (minikube), all dependent services (databases, message queues, provider simulators), sample data, and IDE configuration. Service ownership: each engineering team owning 2-4 microservices with independent deployment capability — teams deploying their services without coordinating with other teams or waiting for a deployment window. From monthly deployment coordination to dozens of independent daily deployments. Preview environments: every pull request automatically deployed to a preview environment with a unique URL — product managers, designers, and QA reviewing changes in a production-equivalent environment before merge. Preview environments automatically destroyed after merge.

Results over 12 months. Summer performance: summer 2025 — zero degradation days (versus 18 in 2024). Peak day handling 112,000 bookings (32 percent above 2024 peak) with average search response time of 650ms (versus 12 seconds at 2024 peak). Booking failure rate maintained at 0.6 percent during peak (versus 6.2 percent in 2024). Estimated recovered revenue: EUR 12.8M (the bookings that would have been lost to performance degradation). Provider reliability: provider overbooking incidents from 340 per month (peak season) to 22 per month — an 94 percent reduction. Monthly maintenance windows eliminated — providers no longer experiencing booking blackouts. Provider satisfaction score improving from 3.2/5.0 to 4.4/5.0 — directly impacting provider retention (provider churn from 18 percent to 11 percent). Mobile: app store rating from 3.6 to 4.3 (iOS) and 3.8 to 4.5 (Android) — rating improvement correlating with approximately 8 percent booking conversion increase on mobile. App release frequency from bi-weekly to weekly (iOS) and twice-weekly (Android). Emergency fix deployment from 5-7 days to 24 hours. Deployment velocity: from monthly monolith deployment to average 34 independent service deployments per day. Lead time from feature complete to production: from 6-12 weeks to 3 days. Change failure rate from 12 percent to 1.8 percent with 30-second automated rollback. Developer productivity: new engineer productive deployment within first week (from first month). Engineering time on infrastructure from 28 percent to 9 percent — 120 engineers recovering approximately 9,100 hours per quarter for feature development. Cost: infrastructure cost per booking from EUR 0.82 to EUR 0.39 — a 52 percent reduction. Annual infrastructure cost from EUR 3.2M to EUR 2.1M despite 35 percent booking volume growth. Seasonal cost variation: winter infrastructure cost 60 percent lower than summer — paying only for capacity used rather than year-round peak provisioning. Financial impact: recovered summer revenue EUR 12.8M; provider retention improvement EUR 4.2M (reduced churn value); mobile conversion improvement EUR 6.8M; infrastructure savings EUR 1.1M; developer productivity EUR 3.4M. Total annual impact: EUR 28.3M. Engagement cost: EUR 420K. Annual platform management: EUR 125K. ROI: 67.4x first year.