Media set compute usage（媒体集计算用量）

Media sets bring a number of advanced, out-of-the-box transformations to the platform. In addition to being triggered via transforms and pipelines, media transformations are also triggered by interacting with media items in the platform, for example, by previewing a media item. Additionally, there is a cost to download or stream the full contents of a media item.

Usage is tracked in units of Foundry compute-seconds. The table below describes each transformation available, with usage rate in terms of compute-seconds per gigabyte processed.

If you have an enterprise contract with Palantir, contact your Palantir representative before proceeding with usage calculations.

Transformations

Usage rate is measured in compute-seconds per GB.

All

Transformation	Usage Rate
Download / stream	2

Images

Transformation	Usage Rate
Generate PDF	40
Resize	40
Resize within bounding box	40
Rotate	40
Adjust contrast	75
Annotate	75
Crop / chip	75
Encryption / decryption	75
Geo tile	75
Grayscale	75
Render DICOM image layer	75
Extract text (OCR)	275
Generate embedding	275
Extract text v2 (OCR)	275
Extract layout aware v2	275

Audio

Transformation	Usage Rate
Chunk	75
Select channel	75
Stream with HLS	75
Transcode	75
Waveform generation	75
Transcription	275

Video

Transformation	Usage Rate
Get timestamps for scene frames	40
Extract audio	75
Extract frames at timestamp	75
Chunk	275
Extract all scene frames	275
Stream with HLS	275
Transcode	275

Documents

Transformation	Usage Rate
Get PDF page dimensions	40
Render page as image	40
Render page as image within bounding box	40
Extract all text (raw)	75
Extract form fields	75
Extract table of contents	75
Extract text from pages to array (raw)	75
Extract text on page (raw)	75
Slice PDF range	75
Extract text (OCR)	275
Extract text from pages to array (OCR)	275
Extract layout aware text (raw / OCR)	275
Extract text using VLM *	275

* This usage is accounted for in addition to VLM usage, which is automatically calculated.

Extract text transformation

Format	Raw text mode	OCR mode	Auto* mode
Text	75	275	275
Markdown/HTML	275	275	275

* Auto mode determines, for each page, whether text can be extracted directly or whether OCR is required.

Spreadsheets

Transformation	Usage Rate
Convert sheet to JSON	275

Using media at scale

Media set limits

• There is no limit to how many media items can be uploaded to a media set.
• Transactional media sets have an upload limit of 10,000 items per transaction.
• Transactionless media sets do not have an item upload limit (since there are no transactions).
• Paths of items in media sets cannot exceed 256 characters. Attempting to add an item with a path longer than 256 characters to a media set will result in a MediaSet:MediaItemPathInvalid error.
• Each media item can have a maximum file size of 50 GB.
• For incremental transforms, limiting the batch size of media set inputs is supported.

Throttling

While it is possible to interact with media sets at scale, such as uploading or transforming millions of media items, error handling should be implemented to ensure pipeline stability. When interacting with media sets at scale, the media set service may throttle by responding with Quality of Service (QoS) errors which have status codes of either 429 or 503. This indicates that the media set service cannot handle the storage or compute load at the moment.

Instead of failing on a QoS error when the service throttles, the client should build an adequate retrying mechanism.

For example, if a client is uploading millions of media items by submitting PUT requests at /{mediaSetRid}/items, the client should retry any media item upload requests that receive a QoS error.

When using the transforms-media library in Python transforms, a retrying mechanism is generally already in place. For example, when uploading media files from a catalog dataset to a media set, the library method provided already has retrying logic to handle QoS errors.

from transforms.api import transform, Input, Output
from transforms.mediasets import MediaSetOutput


@transform(
    input_files=Input('/examples/image_files'),
    output_media_set=MediaSetOutput('/examples/image_media_set'),
    uploaded_media_record=Output('/examples/uploaded_media_record'),
)
def compute(input_files, input_files, output_media_set, uploaded_media_record):
    uploaded_media_items = output_files.put_dataset_files_and_get_dataframe_of_uploads(
        input_files,
        ignore_items_not_matching_schema=False,
        ignore_items_failing_to_convert=False,
    )

    uploaded_media_record.write_dataframe(uploaded_media_items)

---

中文翻译

媒体集计算用量

媒体集（Media set）为平台提供了多种开箱即用的高级转换功能。除了通过转换（transforms）和流水线（pipelines）触发外，媒体转换也可以通过平台内与媒体项的交互来触发，例如预览媒体项。此外，下载或流式传输媒体项的完整内容也会产生相应费用。

用量以 Foundry 计算秒（compute-seconds）为单位进行追踪。下表描述了每种可用的转换，以及按每处理千兆字节计算的计算秒数用量率。

如果您与 Palantir 签订了企业合同，在进行用量计算之前，请联系您的 Palantir 代表。

转换

用量率以每 GB 计算秒数为单位。

全部

转换	用量率
下载/流式传输	2

图像

转换	用量率
生成 PDF	40
调整大小	40
在边界框内调整大小	40
旋转	40
调整对比度	75
标注	75
裁剪/切片	75
加密/解密	75
地理瓦片	75
灰度化	75
渲染 DICOM 图像层	75
提取文本（OCR）	275
生成嵌入	275
提取文本 v2（OCR）	275
提取布局感知 v2	275

音频

转换	用量率
分块	75
选择声道	75
使用 HLS 流式传输	75
转码	75
波形生成	75
转录	275

视频

转换	用量率
获取场景帧的时间戳	40
提取音频	75
提取指定时间戳的帧	75
分块	275
提取所有场景帧	275
使用 HLS 流式传输	275
转码	275

文档

转换	用量率
获取 PDF 页面尺寸	40
将页面渲染为图像	40
在边界框内将页面渲染为图像	40
提取所有文本（原始）	75
提取表单字段	75
提取目录	75
提取页面文本到数组（原始）	75
提取页面文本（原始）	75
切片 PDF 范围	75
提取文本（OCR）	275
提取页面文本到数组（OCR）	275
提取布局感知文本（原始/OCR）	275
使用 VLM 提取文本 *	275

* 此用量是在 VLM 用量之外额外计算的，VLM 用量会自动计算。

提取文本转换

格式	原始文本模式	OCR 模式	自动*模式
文本	75	275	275
Markdown/HTML	275	275	275

* 自动模式会为每个页面判断是直接提取文本还是需要 OCR。

电子表格

转换	用量率
将工作表转换为 JSON	275

大规模使用媒体

媒体集限制

• 媒体集中可上传的媒体项数量没有限制。
• 事务性媒体集（Transactional media sets）每笔事务的上传限制为 10,000 项。
• 非事务性媒体集（Transactionless media sets）没有项上传限制（因为没有事务）。
• 媒体集中项的路径不能超过 256 个字符。尝试向媒体集添加路径超过 256 个字符的项将导致 MediaSet:MediaItemPathInvalid 错误。
• 每个媒体项的最大文件大小为 50 GB。
• 对于增量转换（incremental transforms），支持限制媒体集输入的批处理大小。

限流

虽然可以大规模使用媒体集，例如上传或转换数百万个媒体项，但应实现错误处理以确保流水线稳定性。当大规模使用媒体集时，媒体集服务可能会通过返回服务质量（Quality of Service，QoS）错误来进行限流，这些错误的状态码为 429 或 503。这表明媒体集服务当前无法处理存储或计算负载。

当服务限流时，客户端不应因 QoS 错误而失败，而应构建适当的重试机制。

例如，如果客户端通过向 /{mediaSetRid}/items 提交 PUT 请求来上传数百万个媒体项，则客户端应重试任何收到 QoS 错误的媒体项上传请求。

在 Python 转换中使用 transforms-media 库时，通常已经内置了重试机制。例如，当从目录数据集（catalog dataset）上传媒体文件到媒体集时，提供的库方法已经包含处理 QoS 错误的重试逻辑。

from transforms.api import transform, Input, Output
from transforms.mediasets import MediaSetOutput


@transform(
    input_files=Input('/examples/image_files'),
    output_media_set=MediaSetOutput('/examples/image_media_set'),
    uploaded_media_record=Output('/examples/uploaded_media_record'),
)
def compute(input_files, input_files, output_media_set, uploaded_media_record):
    uploaded_media_items = output_files.put_dataset_files_and_get_dataframe_of_uploads(
        input_files,
        ignore_items_not_matching_schema=False,
        ignore_items_failing_to_convert=False,
    )

    uploaded_media_record.write_dataframe(uploaded_media_items)