TLDR Data 2026-07-20

The sites that block scrapers are exactly the ones data teams need, but most benchmarks ignore those. String's open-source benchmark looks at the hard cases.

The results: String's Web Access API hit 96% success on the hardest targets. Browserbase? 51%. Nimble? 59%. Oxylabs? 69%. Firecrawl? 72%.

Real Amazon prices, live Zillow comps, Indeed postings, if it's online, String will reach it, and we'll do it faster and cheaper than your current vendor.

Fetch, search, map, and browser endpoints. Use an API key or MCP.

Try String for free here

Want to brainstorm your hardest use cases? Book time with us and we'll solve them together

DataFusion Comet speeds Spark reads on Iceberg by keeping Java planning while using Iceberg Rust for native Arrow execution. On a 3 TB TPC-DS benchmark, it accelerated 102 of 103 queries and cut total runtime by about 40%, while its use of Iceberg Java's Spark test suites has helped drive more than 40 Iceberg Rust pull requests.

OpenTelemetry traces can convert production AI-agent feedback into training labels for local 7B-13B coding models. Fine-tuned within hours, they can reach roughly 80-85% of frontier-model quality while lowering inference costs, preserving privacy, and scaling across the enterprise.

Meta compresses its graph of billions of users and advertiser entities into clusters of latent interests, enriches them with LLM-processed ad content, and embeds everything in one shared space where any user or entity can be scored against any other. Learning from broad graph structure instead of only rare conversion events gives the deep funnel a usable signal, and the embeddings feed Meta's ad ranking models like GEM and Andromeda.

Data modeling is still essential, especially in the AI era, because clear structures, standards, and context help teams avoid bad data, wasted work, and constant firefighting. Modern data professionals need a flexible “mixed model” approach that uses the right modeling style for the problem rather than forcing everything into one method.

Enterprise agents are constrained more by legacy infrastructure than model quality. LinkedIn pre-provisioned containers and shifted 80% of orchestration to deterministic code, Walmart governed duplicate internal agents, and Zendesk strengthened data pipelines for its 20B conversations. The lesson is to invest in evals, own the agent harness, and keep workloads portable across models.

Enterprise AI agents need governed context, not just vector search: retrieval can find relevant passages, but it does not preserve source boundaries, business meaning, ownership, or usage restrictions. The proposed approach compiles one governed data-product model into multiple representations: graph, Markdown wiki, portable OKF bundle, and compact sidecars such as TOON/GCF. AI readiness depends on connected, trusted context and repeatable operations rather than catalog presence alone.

Asking a data question is still a ticket, a queue, and a week of bogged down data teams. Cube hands that workflow to AI agents who model, explore, and build reports - working from one governed set of definitions and single source of truth. Brex, Wix, and Patagonia run on Cube. Get started for free

Apache Spark 4.2 adds metric views for defining business metrics, SQL vector similarity search, native geospatial types, and change data capture queries through a new CHANGES clause. Spark Connect and default Arrow optimized Python execution make the engine easier to reach. The release bundles over 1,900 commits from more than 260 contributors.

Microsoft's Ontology Playground is a free, open-source tool for visually exploring, building, importing and sharing ontologies for Microsoft Fabric IQ.

Postgres 19 plans to switch default TOAST compression from pglz to LZ4, while keeping the same unified compression framework across heap storage, TOAST, and B-tree indexes. Variable-length types like TEXT, VARCHAR, BYTEA, and JSONB are compressed automatically via varlena metadata, with rows targeting ~2 KB before spilling out-of-line to TOAST. LZ4 is much faster than pglz in tests, often compresses better, and still fails fast on incompressible data.

Trellis stores agent artifacts, rewards, search data, and causal history as persistent experience graphs accessible through SQL, graph, vector, and temporal queries. Meta found that reusing past experience matched performance about 10x faster and reduced successful-solution token costs by 52%, though excessive memory reuse reduced exploration.

Netflix runs LLM serving end-to-end in-house, using a unified JVM-based platform with gRPC plus an OpenAI-compatible HTTP API, rather than relying on hosted APIs. The stack standardizes deployment through Triton-backed Model Scoring Service, using vLLM as the paved-path engine after re-benchmarking. It also adds custom packaging, zero-downtime rollouts, and patched guided decoding to prevent malformed JSON output.

Airbnb cut LLM evaluation cycles from weeks to under a day by caching generated references and judge scores, since over half of model outputs across candidates were identical strings and judge drift of about 1% per run. Small LoRA adapters with rank under 50 train in under an hour on a single GPU for same day fixes, and a final validation stage runs sampled traffic to catch failures.

Dremio's sale to SAP is presented as evidence that data platforms built around a single lakehouse will struggle with AI's demands for broad reach, freshness and unpredictable querying.

Databricks is reportedly raising $3 billion at a $188 billion valuation after successfully repositioning itself from a data platform into a major enterprise AI provider.