Object Storage for AI Images: Tenant Isolation, Signed URLs and Cleanup

·2 views

Storage costs for AI-generated images are driven mainly by how long files are retained rather than by upload activity. A recommended design keeps originals and thumbnails under tenant-scoped keys, serves them through short-lived signed URLs, and removes temporary derivatives using lifecycle rules. Separate objects for each derivative make regeneration, cache invalidation, and deletion deadlines easier to manage. Because object versioning is absent, overwriting files can destroy data, so content-hash or generation-ID keys and new writes are advised. Lifecycle expiry has a one-day minimum, which suits temporary outputs but not workloads requiring minute-level cleanup.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Healthtech exports: signed URLs and object storage for large, authorized downloads

Large healthtech exports may be generated by one worker and downloaded later by a different session, so the author recommends a delivery design that keeps load off the application. Authentication and authorization stay in the app, and the finished file is written to object storage with a short-lived signed URL given to the authorized browser. Serving bytes from the app server is reserved for small, ephemeral files consumed within one request, since large downloads tie up workers, connections, and local disk. Because a signed URL works as a time-limited capability, it should only be issued after tenant ownership and a completed export state are verified, and object keys should be opaque rather than containing sensitive identifiers. The article treats an oversized-request error as a signal to rethink the delivery boundary rather than raising limits, and notes that traffic patterns and residency rules determine the actual trade-offs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Apache Pinot controller and broker require separate authorization reviews

Apache Pinot's controller manages cluster metadata, tables, and segments, while its broker handles SQL queries and returns data, and the two roles carry different exposure risks. Pinot's documentation makes authentication and ACL configuration independent for each component, and authentication is disabled by default, with the default access factory permitting all access. The article recommends explicitly verifying that authentication, authorization, TLS, and internal service credentials are enabled in production deployments. It also advises testing read-only, table-specific, and administrator accounts separately and restricting network exposure of controller and broker endpoints. Search-engine counts of Pinot instances are described as matching assets rather than evidence that access controls are misconfigured.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to accept Telegram Stars in a bot: the XTR payment flow

Developers selling digital goods in Telegram bots and Mini Apps must use Telegram Stars, with the currency code XTR. The article describes a payment flow that starts by writing a pending order to the database and using its ID as the invoice payload. Invoices are created with sendInvoice in chats or createInvoiceLink in Mini Apps, using an empty provider token and a single whole-number Stars price. On receiving a pre_checkout_query, the server must validate the order within Telegram's 10-second window before any Stars move. The author stresses that a client-reported "paid" status should never grant an item without server-side confirmation, and that duplicate or unexpected charges must be detected and refunded.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Amazon Swaps Frontier LLM for 2B Model in Ticket Routing, Cutting Cost 98%

Amazon reportedly replaced Claude Sonnet 4.6 with a 2-billion-parameter model on one of its large internal ticketing systems, reporting similar quality at a fraction of the cost. Cost per ticket reportedly fell from about $0.0222 to roughly $0.0005. The gains came less from cheaper tokens than from changing the task: the smaller model outputs a single option index rather than generating a label and explanatory text. This also cut latency by about 78% and returned a confidence score that allows uncertain tickets to be flagged. Cost also depends on billing model, since a dedicated instance is cheaper at high volume but carries a fixed hourly floor.

0 comments Read more at DEV Community