September 2026 Summaries

Rising demand for AI infrastructure has increased DRAM costs and constrained supply, creating challenges for organizations that rely on in-memory databases to support growing datasets, real-time decision systems, caches, feature stores, and AI agents. Reducing retained data or shifting more requests to slower systems can lower memory spending but may weaken application performance and decision quality, particularly for high-volume fraud, recommendation, and agentic workloads. The proposed approach is storage tiering, in which frequently accessed data remains in RAM while less active keys and values are stored on lower-cost SSDs and loaded into memory when needed, allowing applications to retain a single database interface. Redis presents Redis Flex as an implementation of this model, offering configurable RAM-to-SSD ratios, claimed lower per-gigabyte costs, no required code changes, and preview support for Redis Search indexes stored on SSD.

Rising demand for AI-oriented high-bandwidth memory has constrained conventional server DRAM supply, contributing to sharp memory price increases that may persist until new manufacturing capacity arrives in 2027. This trend raises costs for in-memory database users, particularly because much stored data is infrequently accessed, while reducing retention, shrinking caches, or distributing data across multiple systems can impair performance and increase operational complexity. The discussion presents RAM-and-SSD tiering as an alternative, emphasizing that systems able to move both keys and values to SSD can reserve RAM for frequently accessed data while retaining less-active data at lower cost. Redis Flex is presented as a tiered Redis option that lets users select a 10% to 50% RAM ratio, stores hot data in RAM and colder keys and values on SSD, supports terabyte-scale datasets and Redis Search indexes, and is claimed to reduce per-gigabyte costs by up to 80% compared with RAM-only deployments without application code changes.

Redis has expanded its strategic collaboration agreement with AWS to support organizations building and scaling real-time and AI-powered applications on AWS. The initiative combines Redis’s data capabilities for caching, vector search, semantic caching, session management, and agent memory with AWS infrastructure and services such as Amazon Bedrock, aiming to help teams move generative AI projects from prototypes into production. The companies will focus on developing AI reference architectures and technical guidance, modernizing performance-sensitive applications, easing migration to Redis Cloud on AWS, and simplifying purchasing through AWS Marketplace, including private offers and enterprise procurement options. Redis and AWS also plan to coordinate sales, partner, and technical teams through customer workshops, account planning, and solution campaigns, building on their existing collaboration around cloud migration, application modernization, and go-to-market activities.

Redis has released the official fastapi-redis-sdk, developed with the FastAPI team to provide Redis integration that follows FastAPI’s dependency-injection and application-lifecycle patterns. The SDK supports synchronous and asynchronous dependencies, connection management, HTTP caching behavior, cache-control headers, and deployment across self-managed FastAPI applications and FastAPI Cloud, where Redis Cloud is available as a native integration during the platform’s public beta. Its caching features allow endpoints to cache GET responses with configurable time-to-live values, invalidate related entries after DELETE operations, update caches following successful PUT operations, and organize related data through eviction groups, while rate-limiting tools support both individual and stacked burst and sustained limits. Configuration uses Pydantic Settings and environment variables or local .env files for portability and type safety. Advanced capabilities include conditional caching, cascading invalidations, dynamic TTLs, downstream-resource protection, per-tenant fairness, and runtime token-budget limits, with documentation, a repository, and a PyPI package available for developers to try.

Redis’s MultiDBClient for Python applications provides client-side geographic failover for Active-Active Redis deployments, where CRDT-based replication keeps data synchronized across regional instances that can each accept reads and writes. The wrapper monitors configured Redis or Redis Cluster endpoints, selects an active database based on health and weights, automatically fails over when an endpoint becomes unhealthy, and can fail back to a preferred recovered endpoint. It combines proactive background health checks with reactive, circuit-breaker-based failure detection, while retries can target the newly active endpoint; replication itself remains handled by the Active-Active Redis layer. Configuration options cover endpoint weights, client types, health-check policies, retry behavior, failure windows, failover timing, and optional manual endpoint selection, allowing teams to balance rapid recovery, tolerance for transient errors, and operational overhead. MultiDBClient also offers event notifications and observability metrics for failover transitions, and Redis recommends tuning settings according to traffic volume, latency requirements, and acceptable false-positive failovers.

Migrating from Azure Cache for Redis to Azure Managed Redis (AMR) may reduce monthly service costs substantially when workloads fit the target configuration, with East US 2 comparisons showing roughly 43–44% savings for high-availability Premium-to-Balanced deployments. For example, a two-node Premium P3 instance costing about $1,619 per month could be replaced by a Balanced B20 configuration at about $918, potentially saving approximately $8,410 annually before migration expenses. Savings vary by existing tier, capacity needs, availability requirements, regional topology, reservations, and additional features, and can disappear if a workload requires a larger or compute-optimized AMR tier. Migration commonly involves updating connection and clustering settings, testing application behavior, planning data transfer or cache rebuilding, and preparing for cutover and rollback, while total-cost calculations should include commercial terms, engineering effort, temporary parallel environments, networking, monitoring, and transfer charges. An illustrative $5,000 P3-to-B20 migration budget would be recovered in about eight months through monthly savings of roughly $701, though higher migration costs extend payback time. With Azure Cache for Redis Enterprise offerings scheduled for retirement in 2027 and Basic, Standard, and Premium tiers in 2028, organizations are encouraged to evaluate workload metrics, sizing, costs, and migration requirements ahead of those deadlines.

Redis introduced Smooth Scaling for Redis Cloud Pro, a capability designed to make database capacity changes faster and less disruptive to live applications while preserving the existing scaling workflow. Built on atomic slot migration from Redis 8.4, it moves only the necessary hash slots directly between shards, reducing unnecessary data movement and avoiding intermediate migration states. This can shorten scaling operations, reduce client redirects, connection timeouts, throughput dips, and tail-latency spikes, although results vary by dataset, workload, and requested capacity change. Smooth Scaling requires Redis version 8.4 or later, the default Redis hashing policy, and a RAM-based database; it does not currently support Active-Active or Redis Flex databases. The feature is rolling out gradually, activates automatically for eligible databases, requires no code or configuration changes, and leaves databases using standard or custom hashing policies unchanged.

Redis announced several enterprise-focused capabilities intended to reduce operational complexity and support growing real-time application and AI workloads across cloud, on-premises, and hybrid environments. Redis Radar provides centralized visibility into Redis deployments, including open-source instances, to aid capacity planning, governance, and resource management, while new Datadog and Grafana integrations connect Redis to existing observability workflows. Enhanced Redis Data Integration supports multiple sources and independently managed pipelines, enabling organizations to synchronize data from diverse systems into a unified Redis layer for faster application and agent access. Smooth Scaling for Redis Cloud Pro improves the efficiency and continuity of user-initiated capacity changes without introducing automatic scaling. Redis also added Search to Redis Flex, which uses RAM for hot data and SSD for larger datasets and indexes, expanding support for large-scale search use cases while reducing dependence on memory capacity alone.

Redis has announced a native Datadog integration for Redis Cloud Pro in Private Preview, allowing customers to stream Redis Cloud metrics directly into Datadog without installing an agent, configuring VPC peering, or managing firewall rules. Users connect the services through the Redis Cloud console by supplying a Datadog API key and selecting the corresponding US or EU region, after which metrics generally appear within 30 seconds under the rdse2 prefix. The integration supports monitoring database availability, memory consumption and limits, key counts, request volumes and latency, active client connections, cache hits and misses, and host memory, with filters for subscriptions and databases. A Redis Cloud Database Dashboard is intended to be automatically available in Datadog, though its rollout may follow the preview launch, and users can manage credentials, regions, connectivity, and enablement from the Redis Cloud console. During Private Preview, access and support are handled through Redis account teams rather than standard support channels.

Redis has introduced Redis Radar, a generally available fleet-management tool that consolidates inventory, health, usage, configuration, alerts, licenses, and certificate information across Redis Software, Redis Cloud, Redis Open Source, Amazon ElastiCache, and Google Memorystore deployments. Available as either a managed service or self-managed software for secure and air-gapped environments, Radar connects to platforms using read-only credentials, APIs, or an outbound TLS-based agent and collects only deployment metadata and metrics rather than application data. By providing a centralized view of versions, capacity, security settings, database status, and impending certificate or license expirations, Radar is intended to help teams identify operational issues, conduct routine reviews, and plan upgrades or capacity needs without manually navigating multiple administrative consoles.

Redis Data Integration (RDI) Software and Redis Cloud now support multi-source pipelines, enabling multiple databases to feed captured and transformed data into a single Redis target database. The capability supports sources including Snowflake, MongoDB, Oracle, MySQL, PostgreSQL, RDS, and Aurora, helping organizations consolidate distributed data across regions, business units, shards, or technology stacks into a unified low-latency layer. Intended use cases include real-time fraud and risk checks, consolidated portfolio views, unified regional or tenant data, and federated-cache architectures, while reducing application-side data stitching and separate integration deployments. Redis also plans to introduce multi-pipeline support, which will allow a single RDI installation to operate multiple pipelines for distinct workloads and ingestion flows, expanding its role as an integration layer for increasingly complex real-time Redis applications.

Redis 8.10, now generally available in Redis Open Source, introduces efficiency, data-structure, query, and operational improvements aimed at large-scale deployments. Compact hash templates allow similarly structured hashes to share field names, reducing memory use by up to 50%, while HIMPORT can double bulk hash-loading throughput; additional optimizations improve hash writes and Stream reads and reduce memory use for deep Streams. Expanded JSONPath supports arithmetic, filtering, string and array functions, and aggregations, while Streams gain MAXCOUNT and MAXSIZE controls for bounding multi-stream read responses. New Set cardinality commands calculate union and difference sizes without retrieving members, Lists gain atomic multi-element movement options, and TimeSeries adds timestamp-grouped multi-series queries, blocking reads for incoming samples, and exclusion of empty results. The release also adds coordinated incremental backup and restore through new BACKUP commands, using staggered base snapshots and incremental AOF files to reduce CPU and memory spikes during Redis Cluster backups.

Redis has introduced a Redis Development plugin for ChatGPT Work and Codex that supplies current engineering guidance directly within AI-assisted development workflows, helping teams build, review, and troubleshoot Redis applications without repeatedly consulting separate documentation. The plugin includes specialized skills covering core data modeling, connections, search and vector retrieval, semantic caching, clustering, security, observability, and Redis Agent Memory, with guidance for newer features such as LangCache and hybrid retrieval pipelines. Separately, OpenAI’s Data agent in ChatGPT Work can connect to approved Redis data sources, allowing users to investigate data and changes through natural-language questions. Redis reports that evaluations across multiple models showed improved pass rates and more direct responses when the skills were enabled. The plugin is available through supported ChatGPT Work and Codex interfaces, while the existing skills CLI remains supported for other agents, and its MIT-licensed repository accepts community contributions and feedback.

Databricks Real-Time Mode (RTM) and Redis are presented as a complementary architecture for applications requiring continuous event processing and near-instant data serving, such as e-commerce recommendations that adapt during a user session. RTM extends Spark Structured Streaming with sub-second processing and p99 latency in the tens to low hundreds of milliseconds, allowing teams to use existing Spark APIs rather than operate a separate streaming engine, while Redis serves computed session state, recommendations, or scores through sub-millisecond reads and writes. In the described implementation, clickstream events enter Kafka, RTM sessionizes user activity and applies recency- and intent-weighted product scores, then a Spark ForeachWriter stores ranked recommendations and product metadata in Redis sorted sets and hashes with short TTLs. A benchmark using Azure Event Hubs and Azure Managed Redis reportedly sustained 100,000 events per second across 10,000 active users, achieving 157 ms p99 click-to-recommendation latency and approximately 530,000 Redis operations per second without evictions. The pattern is also positioned for fraud detection, online ML feature serving, inventory updates, fleet tracking, security operations, dashboards, and multi-agent coordination, with support across Databricks cloud deployments and several Redis management options.