This publish is co-written with Moshiko Ben Abu, Software program Engineer at CyberArk.
CyberArk achieved as much as 95% discount in case decision time utilizing Amazon Bedrock and Apache Iceberg.
This enchancment addresses a problem in technical assist workflow: when a assist engineer receives a brand new buyer case, the largest bottleneck is commonly not diagnosing the issue however getting ready the information. Buyer logs arrive in numerous codecs from a number of distributors, and every new log format sometimes requires guide integration and correlation earlier than an investigation can start. For easy instances, this course of can take hours. For extra advanced investigations, it could take days, slowing decision and decreasing total engineer productiveness.
CyberArk is a world chief in id safety. Centered on clever privilege controls, it offers complete safety for human, machine, and AI identities throughout enterprise purposes, distributed workforces, and hybrid cloud environments.
On this publish, we present you the way CyberArk redesigned their assist operations by combining Iceberg’s clever metadata administration with AI-powered automation from Amazon Bedrock. You’ll discover ways to simplify knowledge processing flows, automate log parsing for various codecs, and construct autonomous investigation workflows that scale mechanically.
To attain these outcomes, CyberArk wanted an answer that might ingest buyer logs, mechanically construction them, set up relationships between associated occasions, and make the whole lot queryable in minutes, not days. The structure needed to be serverless to deal with unpredictable assist volumes, safe sufficient to guard buyer Personally Identifiable Info (PII), and quick sufficient to permit similar day case decision.
The legacy structure: Bottlenecks and guide workflows
When assist engineers obtained buyer instances, they’d add log information to the information lake saved in Amazon Easy Storage Service (Amazon S3). The unique design then suffered from the complexity of multi-step uncooked knowledge processing.
First, CyberArk’s customized parsing logic operating on AWS Fargate would parse these uploaded log information and remodel the uncooked knowledge. Throughout this stage, the system additionally needed to scan for PII and masks delicate knowledge to guard buyer privateness.
Subsequent, a separate course of transformed the processed knowledge into Parquet format.
Lastly, AWS Glue crawlers had been required to find new partitions and replace desk metadata for processed Parquet information. This dependency grew to become essentially the most advanced and time-consuming a part of the pipeline. Crawlers ran as asynchronous batch jobs moderately than in actual time, typically introducing delays of minutes to hours earlier than assist engineers might question the information.
However the inefficiency went deeper than simply architectural complexity. CyberArk helps clients operating various product environments throughout a number of distributors. Every vendor and product produces logs in numerous codecs with distinctive schemas, area names, and buildings. Including assist for a brand new vendor meant days of integration work to know their log format and construct customized parsers.
Determine 1: Legacy log ingestion structure diagram displaying the circulate from S3 add via AWS Fargate processing with AWS Glue Crawler
Past ingestion, the investigation course of itself was guide and time consuming. Assist engineers would manually question knowledge, correlate occasions throughout completely different log sources, search via product documentation, and piece collectively root trigger evaluation via trial and error. This course of required deep product experience and will take hours or days relying on concern complexity. The brand new structure addresses these inefficiencies via three key improvements:
- Single stage serverless processing: AWS Fargate with PyIceberg instantly creates Iceberg tables from uncooked logs in a single cross, eradicating intermediate processing steps and crawler dependencies fully.
- AI powered dynamic parsing: Amazon Bedrock mechanically generates grok patterns for log parsing by analyzing file schemas, remodeling what was as soon as a guide, time consuming course of into a totally automated workflow.
- Autonomous investigation with AI Brokers: AI Brokers autonomously carry out full root trigger evaluation by querying log knowledge, analyzing product data bases, figuring out occasion flows, and recommending options, remodeling hours of guide investigation into minutes of automated intelligence.
The answer: AI-powered automation meets single-stage Iceberg processing
The brand new system delivers zero contact log processing from add to question. Assist engineers merely add buyer log ZIP information to the system. Right here’s the place the transformation occurs: CyberArk’s customized processing logic nonetheless runs on AWS Fargate, however now it makes use of Amazon Bedrock to intelligently perceive the information.
Zero-touch log processing workflow
The system extracts pattern log entries from the uploaded log information and sends them to Amazon Bedrock together with context concerning the log supply and desk schema from AWS Glue Information Catalog. Amazon Bedrock analyzes the samples, understands the construction, and mechanically generates grok patterns optimized for the particular log format.
Grok patterns are structured expressions that outline extract significant fields from unstructured log textual content. For instance, the next grok sample specifies {that a} timestamp seems first, adopted by a severity degree, then a message physique %{TIMESTAMP_ISO8601:timestamp} %{LOGLEVEL:severity} %{GREEDYDATA:message}
The system validates these grok patterns towards extra samples to confirm accuracy earlier than making use of them to parse the entire log file. Efficiently validated grok patterns are saved in Amazon DynamoDB, making a repository of identified patterns. When the system encounters related log codecs in future uploads, it could retrieve these patterns instantly from Amazon DynamoDB, avoiding redundant grok sample technology. Amazon Bedrock processes log samples in real-time with out retaining buyer knowledge or utilizing it for mannequin coaching, sustaining knowledge privateness.
This complete course of invokes Claude 3.7 Sonnet mannequin from Amazon Bedrock and is orchestrated by AWS Fargate duties with retry logic for reliability. The processing makes use of these AI-generated grok patterns to parse the logs and create or replace Iceberg tables utilizing PyIceberg APIs with out human intervention.
This automation lowered logs onboarding time from days to minutes, enabling CyberArk to deal with various buyer environments with out guide intervention.
Determine 2: Log ingestion structure diagram displaying the circulate from S3 add via AWS Fargate processing with Amazon Bedrock integration to Iceberg desk creation
Apache Iceberg: Simplified structure, sooner queries
Iceberg simplified and improved CyberArk’s knowledge lake structure by addressing the 2 main bottlenecks within the legacy system: gradual schema administration and inefficient question efficiency.
Constructed-in schema evolution removes crawler dependency
Within the legacy structure, AWS Glue crawlers grew to become a supply of operational overhead and latency. Even when triggered on demand, crawlers ran as batch jobs over S3 prefixes to find partitions and replace metadata. As knowledge volumes grew and datasets diversified throughout distributors and schemas, groups needed to handle and function a rising variety of crawler jobs. The ensuing delays, typically starting from minutes to hours, slowed knowledge availability and downstream investigation workflows.
Iceberg removes this whole layer of complexity. Iceberg’s clever metadata layer mechanically tracks desk construction, schema modifications, and partition info as knowledge is written. When CyberArk’s processing creates or updates Iceberg tables via PyIceberg, the metadata is up to date immediately and atomically. There’s no ready for crawlers jobs to finish, and no danger of stale metadata. The second knowledge is written, it’s instantly queryable in Amazon Athena.
PyIceberg: Making Iceberg accessible past Apache Spark
Working with Iceberg often concerned Apache Spark and the complexity of distributed knowledge processing. PyIceberg modified that by letting CyberArk create and handle Iceberg tables utilizing a easy Python library. CyberArk’s knowledge engineers might write easy Python code operating on AWS Fargate to create Iceberg tables instantly from parsed logs, with out spinning up Spark clusters.
This accessibility was important for CyberArk’s serverless structure. PyIceberg enabled single stage processing the place AWS Fargate duties might parse logs, apply PII masking, and create Iceberg tables in a single cross. The end result was less complicated code and decrease operational overhead.
Metadata-driven question optimization delivers velocity
Along with eradicating crawlers, Iceberg considerably improved question efficiency via its clever metadata structure. Iceberg maintains detailed statistics about knowledge information, together with min/max values, null counts, and partition info. When assist engineers question knowledge in Athena, Iceberg’s metadata layer helps partition pruning and file skipping, ensuring queries solely learn the particular information containing related knowledge. For CyberArk’s use case, the place tables are partitioned by case ID, this implies a question for a selected assist case solely reads the information for that case, ignoring probably hundreds of irrelevant information. This metadata pushed optimization lowered question execution time from minutes to seconds, permitting assist engineers to interactively discover knowledge moderately than ready for outcomes.
ACID transactions preserve knowledge consistency
In a multi consumer assist atmosphere the place a number of engineers could also be analyzing overlapping instances or importing logs concurrently, knowledge consistency is crucial. Iceberg’s ACID transaction assist helps confirm that concurrent writes don’t corrupt knowledge or create inconsistent states. Every desk replace is atomic, remoted, and sturdy, offering the reliability CyberArk wanted for manufacturing assist operations.
Time journey permits historic evaluation
Iceberg’s built-in versioning permits assist engineers to question historic states of knowledge, important for understanding how buyer points advanced over time. If an engineer must see what the logs regarded like when a case was first opened versus after a buyer utilized a patch, Iceberg’s time journey capabilities make this easy. This characteristic proved important for advanced troubleshooting situations the place understanding the timeline of occasions was vital to decision.
Automated desk optimization with AWS Glue
Iceberg tables require periodic upkeep to keep up question efficiency.
CyberArk enabled AWS Glue computerized desk optimization for his or her Iceberg tables, which handles compaction and expired snapshot cleanup within the background.
For CyberArk’s steady add workflow, this automation avoids efficiency degradation over time. Tables keep optimized with out guide intervention from the engineering crew.
AI Brokers: Autonomous investigation workflow
Whereas the Claude 3.7 Sonnet mannequin from Amazon Bedrock automates grok sample technology for log ingestion, the extra superior use of Amazon Bedrock comes within the investigation workflow. We use AI brokers with Bedrock fashions to alter how assist engineers analyze and resolve buyer points.
From guide evaluation to AI powered investigation
Within the legacy workflow, assist engineers would manually question knowledge, correlate occasions throughout completely different log sources, search via product documentation, and piece collectively root trigger evaluation via trial and error. This course of required deep product experience and will take hours or days relying on concern complexity. AI Brokers automate this whole investigation course of. Assist engineers use an inside portal to ask questions in pure language about buyer points, questions like
“Present me authentication errors for case 12345 within the final 24 hours”, “What had been the most typical errors throughout instances opened this week?” or “Evaluate the error patterns between case 12345 and case 12346.”
Behind the scenes, the system fires specialised AI Brokers that autonomously carry out thorough evaluation.
How assist brokers work
Every AI Agent operates as an clever investigator with a transparent mission: perceive what occurred, decide why it occurred, and advocate repair it. When a assist engineer asks a query, the agent collects related knowledge by querying Athena to retrieve log knowledge from Iceberg tables, filtering for the particular case and time interval related to the investigation. The agent then accesses CyberArk’s inside data base for the particular product concerned, understanding identified points, frequent error patterns, and documented options. The agent then performs the next evaluation:
- Movement identification: Analyzes the sequence of occasions within the logs to know what really occurred in the course of the buyer’s concern
- Root trigger dedication: Correlates log occasions with product data to establish the underlying reason behind the issue
- Answer suggestions: Suggests particular remediation steps based mostly on the basis trigger evaluation and identified decision patterns
This complete course of occurs in minutes, delivering superior evaluation that might have taken assist engineers hours to carry out manually.
For advanced instances the place an answer isn’t discovered, the assist agent escalates to a different, specialised agent that interacts with service engineers to gather extra inputs and experience. This human-in-the-loop strategy makes certain that even essentially the most difficult instances obtain acceptable consideration whereas nonetheless benefiting from the automated investigation workflow. The insights gathered from these escalated instances are mechanically fed again into CyberArk’s data base, constantly bettering the system’s capacity to deal with related points autonomously sooner or later.
Amazon Bedrock by no means shares buyer knowledge with mannequin suppliers or makes use of it to coach basis fashions, case knowledge and investigation insights stay inside CyberArk’s atmosphere.
Concurrent agent execution at scale
When a number of assist engineers examine completely different instances concurrently, the answer runs specialised brokers concurrently. CyberArk at present makes use of Claude 3.7 Sonnet as the muse mannequin for these brokers. Every agent works independently on its assigned investigation, working in parallel with out useful resource rivalry. This concurrent execution permits the investigation workflow to scale mechanically with assist quantity, dealing with peak masses with out efficiency degradation.
AI-powered investigation benefit
This AI-powered investigation workflow delivers two key benefits.
Investigations that took hours now full in minutes, enabling assist engineers to resolve as much as 4x extra instances per day.
The system additionally creates a steady studying suggestions loop. When instances require guide decision by engineers, these resolutions are mechanically recorded and fed again into the data base. Future investigations profit from this collected experience, with brokers making use of classes realized from earlier guide resolutions to related instances. Amazon Bedrock doesn’t use buyer knowledge to coach basis fashions. Case knowledge and investigation insights stay inside CyberArk’s atmosphere.
This automated suggestions mechanism means the investigation workflow turns into more practical over time, constantly bettering decision accuracy and velocity.
Determine 3: Investigation workflow diagram displaying pure language question via AI Brokers to Athena queries and data base evaluation
Scaling with out proportional engineering progress
The enterprise influence of this AI automation is important. CyberArk can develop its vendor protection and product portfolio with out including knowledge engineering headcount. The identical system that handles at this time’s log sorts will mechanically deal with tomorrow’s additions, whether or not that’s ten new codecs or hundreds, considerably decreasing time to marketplace for new product and vendor integrations.
The outcomes: Important enhancements in decision time and productiveness
The transformation delivered measurable enhancements throughout each key metric.
Decision time: CyberArk achieved as much as 95% discount in time from case project to decision. Easy instances that used to take 4 to six hours now take simply 15 to half-hour. Advanced instances that beforehand took as much as 15 days are actually accomplished in 2 to 4 hours.
Engineer productiveness: Assist engineers now deal with 8 to 12 instances per day, in comparison with simply 2 to three instances earlier than. This implies every engineer helps as much as 4x extra clients.
Information availability: Logs are queryable inside minutes of add as a substitute of ready hours or days. Assist engineers can begin investigating points virtually instantly after receiving buyer knowledge.
Operational effectivity: The system requires zero guide intervention for brand spanking new log codecs or schema modifications. Circumstances that used to require days of knowledge engineering work now occur mechanically.
Price optimization: The serverless structure alleviated idle infrastructure prices whereas scaling mechanically with demand. CyberArk solely pays for what they use, after they use it.
Buyer satisfaction: Quicker decision instances and proactive concern identification considerably improved the shopper expertise. Issues get solved in hours as a substitute of days, and clients spend much less time ready for solutions.
What’s subsequent?
Whereas AWS continues to innovate throughout each knowledge lake administration and agentic AI infrastructure, the next capabilities align effectively with CyberArk’s structure and will provide extra operational advantages because the system scale.
Agent infrastructure maturity
Because the agent-based structure scales to deal with hundreds of concurrent investigations, CyberArk is transitioning to Amazon Bedrock AgentCore for future agent deployments. AgentCore offers a managed runtime for manufacturing AI brokers with enhanced observability via AWS X-Ray integration, clever reminiscence for context retention throughout periods, and streamlined operational workflows. Whereas the present AI Brokers implementation delivers the efficiency and reliability CyberArk wants at this time, AgentCore represents a pure evolution path as operational necessities develop, providing framework-agnostic deployment, computerized scaling, and complete monitoring capabilities with out infrastructure administration overhead.
Amazon S3 Tables
CyberArk’s present structure makes use of Iceberg tables saved in Amazon S3 buckets. Amazon S3 Tables provides totally managed Iceberg tables with built-in optimization.
As CyberArk proceed to scale with tons of of Iceberg tables and speedy knowledge progress, CyberArk is exploring a migration to Amazon S3 Tables to additional scale back operational overhead.
S3 Tables take away the necessity to arrange and monitor AWS Glue upkeep jobs. It mechanically performs upkeep to boost the efficiency of Iceberg tables, together with unreferenced file removing, file compaction, and snapshot administration. Moreover, S3 Tables offers Clever-Tiering that mechanically strikes knowledge between storage courses based mostly on entry patterns, optimizing storage prices with out guide intervention.
As a result of S3 Tables makes use of Iceberg open desk format, migration wouldn’t require modifications to present Athena queries and PyIceberg code. This flexibility permits CyberArk to guage and undertake S3 Tables when the operational and price advantages align with their enterprise wants.
Conclusion
CyberArk’s transformation demonstrates how combining trendy knowledge lake structure with AI automation can considerably change operational economics. By combining Iceberg’s clever metadata administration with AI-powered automation from Amazon Bedrock, CyberArk reworked case decision from days to minutes whereas enabling assist operations to scale mechanically with enterprise progress. Assist engineers now spend their time fixing buyer issues as a substitute of wrangling knowledge, clients obtain sooner resolutions, and the system scales mechanically with the enterprise.
To be taught extra about Iceberg on AWS, confer with Working with Amazon S3 Tables and desk buckets and Utilizing Apache Iceberg on AWS. To be taught extra about Amazon Bedrock AgentCore, confer with Amazon Bedrock AgentCore.
Concerning the authors
