Facebook, Inc.

Meta Platforms, Inc.

Hongzhong Jia

Xiaojun Liang

Jason M Taylor

Jay Parikh

Narsing Krishna Vijayrao

Li Hua

Goranka Bjedov

Lin Qiao

Rajiv Krishnamurthy

Aditya Rajkumar Kalro

Andrew Barkett

John Morrow

Daniil Neiter

Domas Mituzas

Lukasz Wesolowski

Joshua David Metzler

Yusuf Abdulghani

Mohamed Fawzi Mokhtar Abd El Aziz

Ming Ni

Mark Douglas Callaghan

A machine learning model is trained. A feature importance metric is determined for each machine learning feature of a plurality of machine learning features of the machine learning model. Based on the feature importance metrics, one or more machine learning features of the plurality of machine learning features of the machine learning model are managed.

Managing machine learning features

In one embodiment, a method includes establishing access to first and second different computing systems. A machine learning model is assigned for training to the first computing system, and the first computing system creates a check-point during training in response to a first predefined triggering event. The check-point may be a record of an execution state in the training of the machine learning model by the first computing system. In response to a second predefined triggering event, the training of the machine learning model on the first computing system is halted, and in response to a third predefined triggering event, the training of the machine learning model is transferred to the second computing system, which continues training the machine learning model starting from the execution state recorded by the check-point.

Distributed training and prediction using elastic resources

A dataset management system ('system') reduces the amount of data to be stored for future analyses. The system determines a sampling rate of the data based on a required level of accuracy, and samples the data at the determined sampling rate. Initially, all data transactions ('full dataset') and the sampled data (&#x201c;sampled dataset&#x201d;) are logged and stored. Based upon a trigger condition, e.g., after a specified period, the full dataset and the sampled dataset are analyzed separately and the analysis results are compared. If the comparison is sufficiently similar (i.e., the sampling produces a sufficiently accurate set of data or a variance between the analysis results of the datasets is within a specified threshold), the system discontinues full data logging and stores only the sampled dataset. Further, the full dataset is deleted. The sampling thus reduces the required data volume significantly, thereby minimizing consumption of the storage space.

Reducing data storage requirements

Embodiments are described for dynamically responding to demand for server computing resources. The embodiments can monitor performance of each of multiple computing systems in a data center, identify a particular computing system of the multiple computing systems for allocation of additional computing power, determine availability of an additional power supply to allocate to the identified computing system, determine availability of a capacity on a power distribution line connected to the particular computing system to provide the additional power supply to the particular computing system, and allocate the additional computing power to the identified computing system as a function of the determined availability of the additional power supply and the determined availability of the capacity on the power distribution line.

Dynamically responding to demand for server computing resources

A chat messaging service provided for a chat user is migrated. At a second chat server from a first chat server, static information associated with a chat user is received. At the second chat server from the first chat server, dynamic information associated with the chat user is received. At least a portion of the dynamic information is received after the chat user is indicated as being associated with the migration state. After the chat user is no longer indicated as being associated with the migration state, a chat message for the chat user is received at the second chat server.

Migrating a chat message service provided by a chat server to a new chat server

The disclosure is directed to receiving a digitized content item that is indicated to be shared with users (e.g., all, some, or one) of a social networking system; selecting, based on one or more characteristics of the digitized content item, a second server computing device at which to further compute or store the digitized content item; and transmitting, to the second computing device, the digitized content item for storage at the second computing device. At least one of the characteristics can be a classification of a user who transmitted the digitized content item via the data communications network. By selecting second server computing devices based on characteristics, data communications network performance can be improved.

Tiered infrastructure for handling data

The disclosure is directed to storing data in different tiers of a database based on the access pattern of the data. Immutable data, e.g., data that does not change or changes less often than a specified threshold, is stored in a first storage tier of the database, and mutable data, e.g., data that changes more often than immutable data, is stored in a second storage tier of the database. The second storage tier of the database is more performant than the first storage tier, e.g., the second storage tier has a higher write endurance and a lower write latency than the first storage tier. All writes to the database are performed at the second storage tier and reads on both storage tiers. The storage tiers are synchronized, e.g., the set of data is copied from the second to the first storage tier based on a trigger, e.g., a specified schedule.

Tiered data storage architecture

Migrating a chat messaging service provided for a chat user is disclosed. At a second chat server from a first chat server, static information associated with a chat user is received. The static information is received before the chat user is indicated as being associated with a migration state. At the second chat server from the first chat server, dynamic information associated with the chat user is received. At least a portion of the dynamic information is received after the chat user is indicated as being associated with the migration state. After the chat user is no longer indicated as being associated with the migration state, a chat message for the chat user is received at the second chat server.

Techniques to adaptively manage service requests within a multi-server system. In one embodiment, a service request and a service rule associated with the service request are received. Data about operating parameters of at least one server in a multi-server system are also received as part of a feedback loop. A response to the service request based on the service rule and the operating parameters is determined. Execution of the service request may be modified according to a tiered service rule based on the at least one server reaching a capacity threshold. The modification includes omitting an action in execution of the service request.

Methods and systems for adaptive capacity management

Channel performance can be improved in a storage device, such as a flash memory or a flash-based solid state drive, by dynamically provisioning available data channels for both write and read operations. In one aspect, a set of available data channels on a storage device is partitioned into a set of write channels and a set of read channels according to a read-to-write ratio. Next, when an incoming data stream of mixed read requests and write requests arrives at the storage device, the allocated read channels process the read requests on a first group of memory blocks, which does not include garbage collection and write amplification on the first group of memory blocks. In parallel, the allocated write channels process the write requests on a second group of memory blocks, which does include garbage collection and write amplification on the second group of memory blocks.

Allocation of read/write channels for storage devices

Storing data in a cache is disclosed. It is determined that a data record is not stored in a cache. A random value is generated using a threshold value. It is determined whether to store the data record in the cache based at least in part on the generated random value.

Statistical cache promotion

Various embodiments of the present disclosure relate to a cache stickiness index for providing measurable metrics associated with caches of a content delivery networking system. In one embodiment, a method for generating a cache stickiness index, including a cluster stickiness index and a region stickiness index, is disclosed. In embodiments, the cluster stickiness index is generated by comparing cache keys shared among a plurality of front-end clusters. In embodiments, the region stickiness index is generated by comparing cache keys shared among a plurality of data centers. In one embodiment, a system comprising means for generating a stickiness index is disclosed.

Growing community of inventors

Cupertino, CA, United States of America

Hongzhong Jia