Adds a configuration parameter called "slow" to database config which
adds a hardcoded delay 70ms to database operations. This is useful
when testing for performance issues.
This commit removes debug logs that meet these criteria:
- used console.log instead of the log service
- were left in the source tree accidentally
- show up at boot or during stat/readdir operations
This is not a comprehensrve removal; this is a first pass to get a
few of the trivial cases.
- Introduced `input_image` and `input_image_mime_type` as optional parameters in the AIInterfaceService schema to enhance image processing capabilities.
- Updated `GeminiImageGenerationService` to differentiate between data URLs and web URLs for image handling.
- Added coercion support in CoercionService for converting data URLs to streams, including extraction of content type.
- Documented the txt2img method in AI.js to clarify usage for text and image-to-image generation.
Usually Puter has a UID by this point, but if it ever doesn't we can
fall back on using the path with a `LIKE ? + '%'` query. However,
users before a particular date might have fsentries without a path,
so we need to perform a check on the user's timestamp to avoid issues.
is_empty wasn't working correctly while running concurrently with
fetchEntry because because the check in fetchIsEmpty was too
specific. This commit fixes this issue while also adding support for
a 'path' parameter to is_empty.
- Enhanced the `generate` method to accept base64 encoded input images and their MIME types.
- Implemented validation for input image and MIME type to ensure proper usage.
- Updated the content construction logic to handle both text-to-image and image-to-image generation scenarios.
Earlier I actually thought this wasn't running in prod because when I
run Puter locally the "refresh app cache" message doesn't appear.
I ensured that minlog was off but I forgot that around the same time
I also disabled the "tick" log level by default. If it weren't for
that I would've realized this an hour ago. That said, it's kind of
odd that this seems to be the source of the leak, because this isn't
new behavior.
There used to be an app cache here that would be invalidated but I
think this may have been removed because the invalidation logic
was causing issues. A 30 second cache seems reasonable though, so
let's do that for now.
Previously ES6 Module extensions would only work on versions of node
with --experimental-require-module enabled by default, and provided the
entry file had no top-level await. This commit uses "await import"
conditionally if the "type" in package.json is set to "module".
Sometimes the same fs entry is fetched multiple times concurrently
because of parallel methods with separation of concerns.
After this commit, only one parallel fetch will go through and the
others will wait for it. Subsequent fetches will not occur is the
data that was needed is already there.
The 'quick_check' increases the amount of time it takes to perform a
stat operation significantly proportional to the time it takes to do
a stat because an fsentry fetch occurs before a transient
memoization map is available for an operation.
For instances where svc_fs.node was called and explicitly passed a
NodePathSelector (instead of being passed a legacy-style object
with a path attribute) this code was failing. This commit moves the
affected code snippet after input coercion and updates references
to use the coerced input.
style: formatting changes were applied to the code. Instead of using
nested conditional blocks we use labelled breaks
Not sure when this went away, but we were always getting the
permission though the file path instead of ID/UUID from what it
looks like, which was hiding this problem.
