Glue jobs and library to manage conversion of AWS Service Logs into Athena-friendly formats.
Note that this library is under active development. If you run into issues, please file an issue or reach out to @dacort.
AWS Service Logs come in all different formats. Ideally they could all be queried in place by Athena and, while some can, for cost and performance reasons it can be better to convert the logs into partitioned Parquet files.
The general approach is that for any given type of service log, we have Glue Jobs that can do the following:
- Create source tables in the Data Catalog
- Create destination tables in the Data Catalog
- Know how to convert the source data to partitioned, Parquet files
- Maintain new partitions for both tables
The following AWS Service Logs are currently supported:
- Application Load Balancer (ALB)
- ELB (Classic)
- CloudTrail
- CloudFront
- S3 Access
- VPC Flow
The implementation aims to be as simple as possible. Utilizing AWS Glue's ability to include Python libraries from S3, an example job for converting S3 Access logs is as simple as this:
from athena_glue_service_logs.job import JobRunner
job_run = JobRunner(service_name='s3_access')
job_run.convert_and_partition()
That said, if there are additional changes you wish to make to the data, you're likely better off writing your own Glue script.
When creating a job, you must specify a service name - this service name indicates which script to copy over for the Glue job. Here's a mapping of Service Log types to their names.
Service Log | Service Name Keyword |
---|---|
ALB | alb |
ELB | elb |
CloudFront | cloudfront |
CloudTrail | cloudtrail |
S3 Access | s3_access |
VPC Flow | vpc_flow |
Pre-requisite: The AWSGlueServiceRoleDefault
IAM role must already exist in the account. (more details)
Additionally, the Glue script must be deployed in a bucket in the same region as your Glue job. For this reason, publicly available versions of the scripts won't be available just yet but perhaps someday.
๐ฅ Warning These jobs can incur a significant amount of cost depending on how much data you have in your source location!
As an example, a Glue Job to convert 10GB of data takes about 18 minutes to run with 10 DPU - that would only be about $1.32 with current Glue pricing, but many people have a lot of log data. :)
Each log type requires an initial Glue script to bootstrap the job and S3 locations will be different depending on the type of service log. These details are specified below in the Service Log Specifics section.
For now, you'll have to run the created jobs manually to trigger the conversion, but you can create a schedule for it in the Glue console. The Glue job will create the raw and converted tables and containing databases, if necessary, and run the conversion proccess over data found in the source location.
Note: Requires the jq utility
There is an example config file in scripts/example_glue_jobs.json. Copy that to scripts/glue_jobs.json
and modify the service definitions for your desired service logs. Note that the values in the defaults
key can be overridden for each service.
To deploy the scripts to your own bucket, override the RELEASE_BUCKET
environment variable and run the following make commands. The scripts will be packaged up, uploaded to your desired bucket and a Glue job will be created for the specified SERVICE_NAME
.
RELEASE_BUCKET=<YOUR_BUCKET_NAME> make private_release
RELEASE_BUCKET=<YOUR_BUCKET_NAME> make create_job service=<SERVICE_NAME>
SERVICE_NAME
is the key in the config file. Don't change those keys as they are also references to the actual Glue scripts.
Depending on what you put for the JOB_NAME_BASE
variable in the config file, this will create a Glue Job using the latest development branch of this code with a name like JOB_NAME_BASE_LogMaster_ScriptVersion
.
You must deploy the Python module and sample jobs to an S3 bucket - you can use make private_release
as noted above to do so, or make package
and copy both dist/athena_glue_converter_<SCRIPT_VERSION>.zip
and scripts/*
to an S3 bucket.
Glue Jobs for each service log type can be created using an AWS CLI command. Below is an example command for doing so with S3 Access Logs. Note that --extra-py-files
and ScriptLocation
must point to wherever you copied them to.
aws glue create-job --name S3AccessLogConvertor \
--description "Convert and partition S3 Access logs" \
--role AWSGlueServiceRoleDefault \
--command Name=glueetl,ScriptLocation=s3://aws-emr-bda-public/agsl/glue_scripts/sample_s3_access_job.py \
--default-arguments '{
"--extra-py-files":"s3://aws-emr-bda-public/agsl/glue_scripts/athena_glue_converter_latest.zip",
"--job-bookmark-option":"job-bookmark-enable",
"--raw_database_name":"aws_service_logs",
"--raw_table_name":"s3_access_raw",
"--converted_database_name":"aws_service_logs",
"--converted_table_name":"s3_access_optimized",
"--TempDir":"s3://<bucket>/tmp",
"--s3_converted_target":"s3://<bucket>/converted/s3_access",
"--s3_source_location":"s3://<bucket>/s3_access/"
}'
For ALB logs, configure your source location to be the prefix before the logs split out into regions. For example, this usually looks like:
s3://<BUCKET>/<PREFIX>/AWSLogs/<AWS_ACCOUNT_ID>/elasticloadbalancing/
The same holds true for ELB logs as ALB logs. Make sure your source location is configured as above.
Similar to ALB logs, CloudTrail appends it's own prefix. You must provide the prefix to where there regions are located. Similar to:
s3://<BUCKET>/<PREFIX>/AWSLogs/<AWS_ACCOUNT_ID>/CloudTrail/
The other thing to note about CloudTrail logs is that there are a few fields that can contain relatively arbitrary JSON. Due to the complexity of detecting a useable schema from these fields, we have decided to leave them as raw JSON strings.
They can be queried using JSON extract functionality in Athena. In addition, these fields are prefixed with json_
to indicate this characteristic:
- json_useridentity
- json_requestparameters
- json_responseelements
- json_additionaleventdata
Finally, because of how Glue categorizes the CloudTrail data, the "raw" table will not be queryable in Athena.
CloudFront logs are delivered to a single prefix in S3. As such, queries over the "raw" table can timeout of there is a large amount of data in that prefix.
The source location should point to that prefix:
s3://<BUCKET>/<PREFIX>/
S3 Access logs are similar to CloudFront in that they are delivered in a single prefix. Make sure your source location is configured as above.
VPC Flow logs also deliver into the AWSLogs prefix:
s3://<BUCKET>/<PREFIX>/AWSLogs/<AWS_ACCOUNT_ID>/vpcflowlogs/
No - there are a couple reasons for this currently.
- We cannot explicitly define table names with crawlers. For UX simplicity, we wanted to be able to specify table names.
- We already know the schema and the partition layout of the source data.
That depends! Which logs would you like to see? Other log formats we're thinking about include:
- VPC Flow Logs
- AWS Cost and Usage Reports
- Database Audit Logs
Create an issue if you have requests for others!
- Graceful schema upgrades
- Allow for custom backfill timeframe
- Data type conversions when writing to Parquet