Key Takeaways
- Tigris Data provides global, S3-compatible storage specifically for AI companies.
- AI and ML workloads demand specialized, distributed storage beyond traditional cloud offerings.
- Operating independent data centers requires deep expertise in hardware and capacity planning.
- Tigris Data's architecture eliminates egress fees, optimizing costs for large data users.
- AI tools significantly boost developer productivity and shift required engineering skill sets.
Deep Dive
- Tigris Data, an infrastructure company co-founded by Ovais Tariq, provides storage solutions for AI companies with significant data needs.
- Tariq's experience running global storage infrastructure for Uber informed the development of Tigris.
- Tigris competes with foundational cloud storage like S3, building its own hardware and software layers for cost efficiency.
- AI/ML workloads, distributed across specialized cloud providers, require a new storage layer compatible with non-traditional cloud environments.
- New specialized cloud providers for inference or training often lack the extensive service offerings of traditional clouds like AWS.
- AI training workloads involve scanning billions of small files, a pattern where traditional object storage performs poorly.
- Audio AI applications demand low-latency storage close to the user, underscoring the need for distributed solutions.
- Tigris Data employs an immutable, append-only, log-based design, enabling instant, zero-copy snapshots and rollback capabilities.
- Capacity planning for such systems requires considering not just storage space but also IOPS, crucial for high-request AI training workloads.
- Managing 100 petabytes of data, as seen at Uber, involves continuous work with vendors on newer, cheaper SSD technologies like TLCs and QLCs.
- Data center build-outs involve critical power density discussions for cooling and power, even for storage, not just GPUs.
- Operating independent data centers entails managing hardware failures and performing capacity planning, unlike cloud services.
- Networking in co-located, multi-tenant environments is complex, utilizing a combination of AnyCast and GeoDNS for management.
- Tigris Data uses FoundationDB, also employed by Apple for iCloud, to ensure strong consistency and scalability for metadata.
- The guest CEO remains actively involved in coding, developing new features like a caching layer, emphasizing hands-on work in hard tech.
- AI coding tools like GitHub Copilot significantly boost productivity for experienced senior infrastructure engineers.
- AI tools are diminishing the traditional advantage of new college graduates by shifting industry focus to systems thinking and architectural design.
- Some senior developers report up to 80% of their code is AI-generated, enhancing overall productivity and potentially leveling the playing field for junior engineers.
- Tigris Data eliminates egress fees, a significant cost for cloud storage users that can comprise up to 80% of a storage bill.
- The company offers cost advantages through lower operating margins and optimized, modern architectures, unlike legacy cloud providers.
- Starting a new system from scratch allows for cleaner design, specifically benefiting AI workloads and guiding architectural decisions.
- The guest predicts continued growth for specialized service providers in compute, storage, and higher-level services, driven by AI and distributed computing.