Oliver Mannion  

@tweeshan

Software, machine learning and people. Good at reading your tweets. Likes are bookmarks. 🥜🍄🥚

Melbourne, Victoria
Vrijeme pridruživanja: svibanj 2009.

Tweetovi

Blokirali ste korisnika/cu @tweeshan

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @tweeshan

  1. Prikvačeni tweet
    10. lis 2019.

    How we built DeepMatch, a serverless event-driven ML service with a feature serving store

    Poništi
  2. 1. velj

    The first deep-learning project on uses Argo CD for data acquisition, training and inference pipelines

    Poništi
  3. 27. sij

    A 17x decrease in BERT inference latency on CPU (3x on GPU) makes the ONNX runtime worth looking into

    Poništi
  4. 26. sij

    ECS in the enterprise begets a PaaS or at least some tooling. But unlike Kubernetes, I haven't heard as much about ECS platforms. Empire is an open-source PaaS built on ECS & supports a subset of the Heroku API. Probably not the first or last ECS PaaS.

    Poništi
  5. 26. sij

    The Kafka dev experience wasn't fun in 's first approach to Kafka as a service. So they built out a new producer abstraction, sidecar, relay and discovery service. Flink is used on the consumer side. Good to hear these stories 👏

    Poništi
  6. 26. sij

    Sometimes knowing the confidence of a prediction is as important, or more, than being accurate. Hard tho if it's an inverse relationship!

    Poništi
  7. 25. sij

    "The most challenging thing is acquiring patience about how ML-based data products are adopted in the company." 💯

    Poništi
  8. 23. sij

    > I used to be all about “the best language for the problem.” Now I recommend “the language your team knows best, as long as it’s good enough.”

    Poništi
  9. 19. sij

    I'm wondering what the workflow is when you start with interactive development in JupyterHub. Once you've imported dagstermill, can you still run the notebook interactively in JupyterHub? Or must you use dagit?

    Prikaži ovu nit
    Poništi
  10. 19. sij

    Papermill allows you to parameterize Jupyter notebooks and execute them. Dagstermill builds on this by integrating notebooks as a step in a data pipeline. 😎 Inputs, outputs are handled by dagster. Logs and the notebook as executed is stored by dagit.

    Prikaži ovu nit
    Poništi
  11. 19. sij
    Poništi
  12. 18. sij

    This is pretty significant for privacy. A corpus of 3 billion photos (with names?) so presumably they have Facebook photo data. It's hard but legal to scrape LinkedIn. Twitter too I assume.

    Poništi
  13. 18. sij

    The Rust ecosystem is more of a house of cards than I thought (although still better than some other languages)

    Poništi
  14. 12. sij

    Really enjoyed this about WePay's data infra evolution. It's a comprehensive blueprint of problems and solutions at each stage from direct prod database access -> self-serve.

    Poništi
  15. 11. sij

    "How to build a PaaS for 1500 engineers" by * main value-add: integration * don't compete with commercial companies * when components change users shouldn't notice * small automations at scale add up * north star: successful deployments per week

    Poništi
  16. 11. sij

    ING WBBA team's Data Analytics Platform. Notable for: * mostly open-source components * access only via remote desktop * k8s running both interactive & batch workloads (via multi-project quotas + pod priority & pre-emption) * Amundsen + Apache Atlas

    Poništi
  17. 11. sij

    I've created and used overly granular and constraining types before. Knowing when not to is hard.

    Poništi
  18. 10. sij

    A privacy-preserving search engine to compete with Google, using query logs for the *main index* and GBDT for ranking 😮 have open-sourced 😎 * Keyvi, an FST-based key-value store for approximate matching * Granne, graph-based ANN search

    Poništi
  19. 5. sij

    I wonder at what scale focusing on utilization makes the most sense.

    Prikaži ovu nit
    Poništi
  20. 5. sij

    Alibaba moved from multiple clusters to a centralised k8s cluster to maximise utilization. Combining online and office workloads they can get up to 40% CPU utilization. They've a cluster running 10k nodes, which is a big blast radius.

    Prikaži ovu nit
    Poništi
  21. 4. sij

    A great overview of CUE (aka cuelang). compares it to other config tools, describes what makes it unique, and explains how it tackles the challenge of configuration at scale

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·