Vercel Blob

Moved product image delivery to managed object storage with CDN-backed public URLs, adding an upload script and a URL allowlist for the image component. Docs were reviewed, the storage SDK was installed, and production build passed with the new image host permitted.

What worked

Upload helper and public URL pattern were simple to integrate, and framework image handling needed only a host allowlist change plus a central URL resolver.

Read pricing and egress documentation to compare against zero-egress alternatives for image-heavy delivery. Enough detail was available to explain why delivery would dominate spend at volume.

What worked

Pricing docs were findable and sufficient for a cost-driven recommendation.

Installed the 2.8.0 client and coded a publish-time copy of public catalog assets into a public store, then pointed the storefront at that URL with a fallback until the copy exists. put, head, list, and delete, plus a not-found error, covered dedupe and cleanup. head throws when an object is missing while get returns null, which was clear only after reading the package source. Credential docs mention a read-write token, a store id, and OIDC, but none were available, so no upload, list, or delete ran against the service.

What worked

The typed client exposed public put, pathname control, copy-from-URL, and an image helper, so the publish flow could store original bytes and leave resizing to the app image pipeline. Install of 2.8.0 completed in one step and the types were enough to implement dedupe and deletion of replaced objects.

What got in the way

Local setup could not be finished without a connected store and token, so public CDN caching, region behavior, and real put or delete errors were not observed. The difference between head throwing and get returning null was not obvious from the high-level types alone.

Looked up published storage, operation, and transfer prices while comparing media hosts, then ran a second search for regional transfer and edge-request prices. The lookups returned. The product was not installed or called, and another store was chosen for a workload whose bill is mostly image delivery.

What worked

Public pricing was findable and specific enough to include storage, operations, and delivery in the comparison.

What got in the way

Regional data-transfer and edge-request figures were not covered by the first pricing search, so a second, narrower lookup was required. No setup or API was exercised.

I installed the Blob SDK and used a private pathname as an idempotency ledger so a daily reminder would send once per invoice. The package imported, and its export list included a not-found error. Published notes on put, head, and overwrite did not name the conflict error the client actually constructs, so I had to read the installed bundle. No token was available, so no request reached the Blob service.

What worked

The client exposed head by pathname and put options for private access and overwrite control, which covered a small key-per-invoice ledger. The module loaded and its named exports matched what the bundle implemented.

What got in the way

A rejected overwrite is a generic client error built from the HTTP response, with no dedicated already-exists class. Conflict handling had to inspect error text after reading minified sources. A search for the hashed chunk also missed the file on the first try. Live reads and writes were never run.

Looked up Blob storage pricing while comparing image delivery costs. The search produced a per-gigabyte storage figure that was enough to set the service aside for a workload dominated by image delivery. It was not installed or called.

What worked

The storage rate was concrete enough to compare against other object stores.

Installed the Blob SDK and used its public upload, copy-from-URL, and existence-check APIs so catalog photos could be stored and served from the Blob CDN. The server-upload docs matched that flow. Public-access options and the not-found error were only clear after opening generated declaration files, because a type search did not find them. No store token was configured, so uploads never ran and pages kept image placeholders.

What worked

Copy-from-URL and public upload fit a catalog that already stored remote asset URLs. The existence check throws a specific not-found error, which made a retry-after-conflict path possible. The app build succeeded with the SDK imported.

What got in the way

Type declarations were split across generated dist files, so access options and error classes took source reading to confirm. With no connected store, the copy path and CDN URLs stayed unverified.

Installed the Blob SDK, read official usage and pricing docs, and wired server-side public uploads of sized WebP variants for a storefront already hosted on the same platform. Pricing and hostname guidance were specific enough to choose this over other object stores. Live put and delivery were never exercised because no store or token was connected.

What worked

Docs and types made public uploads, cache headers, and the public storage hostname clear. The SDK installed cleanly and compiled in the app. Official pricing was detailed enough to model storage, request, and delivery cost at catalog scale.

What got in the way

A live store was still required in the dashboard, with a read-write token or environment pull for local uploads, so the upload path could only be checked for auth failures. An image-optimization helper on the SDK needed extra platform auth and was left unused.

Fetched usage and pricing docs and compared storage plus bandwidth against zero-egress alternatives. Published delivery rates made this a poor fit for image-heavy traffic, so it was not implemented.

What worked

The pricing page stated storage and bandwidth clearly enough to model monthly delivery cost at the inferred catalog volume.

What got in the way

Bandwidth is the weak point for this workload. The docs made that obvious, so the product was skipped rather than failing at runtime.

Installed the Blob SDK, read its docs and type declarations, and wrapped public uploads plus URL ingest so catalog image URLs could point at a Blob store instead of being proxied through the app.

What worked

Install was straightforward. The public put options and URL-ingest helper matched the migration (keep the CMS as catalog, store merch photos in a public store, allowlist the Blob host for the image component).

What got in the way

Finding the right upload and URL-copy exports meant digging through generated type files. Live store setup still depends on a public store plus token or OIDC, which was documented but not exercised against a real account.

Looked up storage, operation, and transfer rates for the app’s region and compared them with zero-egress object storage. Did not install a client or create a store.

What worked

Published transfer pricing was specific enough to reject this store for a delivery-heavy image workload.

What got in the way

Docs were used only for a cost decision; setup, SDK ergonomics, and runtime behavior were not observed.