Official account of the Xiaohongshu tech team, sharing tech innovations and problem insights, advancing together.
By replacing part of its Spark‑based offline pipeline with StarRocks, Xiaohongshu’s data‑warehouse team cut job execution from hours to minutes, reduced resource usage over 90 %, lowered back‑fill cost by 99 %, and accelerated daily data production by 1.5 hours.
Xiaohongshu’s data‑warehouse team integrated Apache Iceberg‑based data‑lake techniques into its existing warehouse, replacing the legacy Hive/Spark stack with global sorting, Z‑order, and upsert‑enabled tables, which cut query latency by up to 90 %, boosted data freshness by 50 %, slashed storage costs by 83 % and saved tens of thousands of GB‑hours of compute daily.
InstantID, an open‑source tool released by Xiaohongshu in early 2024, generates multiple stylized portraits that preserve a person’s facial identity from a single reference photo in seconds, eliminating fine‑tuning, large storage needs, and multi‑image requirements while seamlessly working with popular diffusion models like Stable Diffusion 1.5 and SDXL.
InstantID is a plug‑in adapter that adds identity‑preserving capabilities to text‑to‑image diffusion models, allowing users to upload a single photo and, within 30 seconds, produce a Spring Festival‑styled portrait with accurate facial features, customizable prompts, and strong text control.
The Encoding‑Alignment‑Interaction (EAI) framework predicts full‑body human motion—including detailed hand joints—by extracting spatio‑temporal features with DCT and GCNs, aligning heterogeneous body‑hand representations via Cross‑Context Alignment, and modeling semantic and physical interactions through Cross‑Context Interaction, achieving state‑of‑the‑art accuracy on the GRAB dataset.
The Controllable Mind Visual Diffusion Model (CMVDM) decodes fMRI signals into semantic vectors and silhouette maps, feeds them into a latent diffusion framework with a ControlNet‑style encoder, and reconstructs high‑fidelity images that surpass existing baselines in both structural similarity and semantic accuracy across multiple brain‑imaging datasets.
Xiaohongshu’s recommendation system uses large‑scale multimodal embeddings, dual‑tower and graph models, and diversity techniques like DPP and SSD to quickly surface high‑quality user‑generated content, enabling ordinary users to gain visibility while balancing personalization, exploration, and efficient LLM‑augmented pipelines.
In a fast‑changing workplace, engineers must pursue self‑success, cooperation and empowerment by developing small and big T‑shaped skills, mastering clear goal‑setting, meeting structures, and the WWH perspective, while leveraging AIGC as a tool, to continuously iterate toward leadership and meaningful career growth.
The AAAI‑2024 paper introduces a Negative Sample Assisted Distillation framework—comprising Negative Assistance Training, Negative Calibration Enhancement, and Adaptive Self‑Consistency—that leverages both correct and incorrect reasoning examples to train a compact LLaMA‑7B student, achieving up to 75.75 % accuracy gains over fine‑tuning on MATH and improving out‑of‑domain benchmarks.
Xiaohongshu’s observability team rebuilt its Prometheus‑based metrics platform using vmagent, dual‑active HA clusters, query push‑down, high‑cardinality governance and multi‑cloud active‑active design, delivering ten‑fold collection speed, up to 70× query capacity, massive CPU‑memory‑storage savings and fully automated scaling.
