• Thursday

  Nov 30, 2023

  • Private work.
  • re:Invent.
    • Watched Swami’s keynote.
      • New: Bedrock support for a few new models. Claude 2.1. Llama 2 70B.
      • Titan text embeddings – converting text to numbers so you can train more efficiently on it.
      • New: multimodal embeddings. Text, image, etc.
      • New: Titan image generation. And they contain invisible watermarks for unique identification!
      • Guest: Nyung Ho, vp ai from intuit. UI, UX, assistants, chat, infra, the expected.
      • Vector databases for bedrock: opensearch, redis, pinecone.
      • New: auroro, mongo, dynamo, documentdb vector support.
      • New: Sagemaker Hyperpod. Train models faster.
      • Guest: Aravind Srinivas, cofounder/ceo perplexity.
      • New: Neptune analytics.
      • New: Zero-ETL integration between opensearch and s3.
      • Guest: Rob Francis, cto booking.com.
      • Q can generate SQL for redshift from natural language. And it can create integrations with AWS glue (ETL).
      • Guest: Shannon Kalisky, pm on quicksight.
    • Lambda docs.
      • IaC: https://docs.aws.amazon.com/lambda/latest/dg/foundation-iac.html.
      • Layers for dependencies/composition: https://docs.aws.amazon.com/lambda/latest/dg/chapter-layers.html.
      • You can also zip dependencies into the artifacts: https://docs.aws.amazon.com/lambda/latest/dg/python-package.html.
      • Testing: https://docs.aws.amazon.com/lambda/latest/dg/testing-guide.html. It’s pretty similar. Test with mocks, test with full cloud (dev) services, etc.
      • Access to other aws services (rds, sqs, etc): https://docs.aws.amazon.com/lambda/latest/dg/access-control-resource-based.html. Basically just adding the IAM/policy for access, then using the language specific sdk for implementation. Examples: boto3, redis-py (or via flask-caching), psycopg2 (or via flask-sqlalchemy), whatever.
      • boto3 is the python sdk (for aws cli): https://boto3.amazonaws.com/v1/documentation/api/latest/index.html.
    • Went through the lambda dependencies rigmarole.
      • You can dynamically install at runtime. Easy, but slows down runtime, obviously (adding costs, hurting perf).
      • You’re supposed to be able to add a requirements.txt to your lambda proj root, and it will autodetect and install. I’ve never seen this work.
      • So you have the two main options remaining.
      • The first is to zip up a hydrated environment and deploy that. Good if you have local testing + CICD set up. It can be folder install or venv. But it’s fast, because all the installed files / deps are already available at init and can be imported by your actual lambda function. But then you can’t use the online editor, etc.
      • Best to go with layers. Same as above. Create a new layer, upload your zip (of the venv’s site-packages folder, with your deps installed, and name the venv “python” so the path is at python/lib/python3.x/site-packages), and then simply add it as a layer in your actual lambda function.
      • Also be wary about macbook vs desktop. The lambda can only run in one architecture, x86 or arm. If you’re building the layer on a DIFF arch than the lambda runtime, it won’t work (psycopg, specifically).
    • Other.
      • Quicksight is their BI tool (vs tableau).
      • AWS cloudformation was released in 2011, SAM in 2016, CDK in 2019.
      • X-ray is the distributed tracing tool.
      • Read aws’ serverless guide: https://docs.aws.amazon.com/serverless/latest/devguide/welcome.html.
      • Esp this page: https://docs.aws.amazon.com/serverless/latest/devguide/starter-lambda.html.
      • Q comes with the aws vscode extension now. You have to create an AWS Builder account (it’s associated with your main aws account, if desired).
  • Probably the hardest spelling bee I’ve seen today?
  • The village for work. Walked past Charlemagne doing an interview in wsp.
  • Instacart hadn’t replied yet, so I submitted the disputes through citi. Wasn’t great over there either. Their interface is finicky – went unavailable multiple times while submitting the 7 disputes. Overall chase and ubereats are significantly better. It’s not close.
  • Final day of sinquefield. Fabi wins. Hans wins tournament of peace with 8/9.
  • Icloud photos are STILL syncing.
  • Supercontest.
    • Pinpoint replied (after 2 days) and said they needed (1) company name (2) tollfree phone number. Provided both.
    • Lots of work on connecting the new lambdas to various AWS resources.
    • Lambda must be in same vpc as rds (and only choose same AZ to avoid cross-AZ charges). Then “connect lambda” on rds side (technically this is just adding the security group, which is the outbound rule lambda->rds over the postgres port).
    • Created a test lambda to query RDS. Associated it with my VPC, and only us-west-1a AZ (same as RDS).
    • Bunch in the prod app due to the final caching deployment sunday night.
      • Harner and I loaded next to each other and he saw my picks on his account. Trying to submit other picks errored with “incorrect CSRF token” exceptions.
      • Good. The view was being cached for the matchups page! That page is specific to user. I’ll unique-ify it by user later during the react change.
      • Removed the view cache for matchups. Checked all other views in the cache, they’re good (not unique by user).