{json} logging platform
at scale
aws athena vs Elasticsearch
Florian DAMBRINE - Senior DevOps Engineer
{"codes": [{"code": "SUCCESS", "ver": "3.0", "company": "gumgum-sports", "labels": [], "annotator": "machine", "server": "TeamClassifier", "date": "20170203125137", "tstamps": [], "id": "TeamClassifier_20170203125137"}], "url_original": "https://www.facebook.com/NYKnicks/photos/a.71815779615.98182.21410634615/10157985435429616/", "url": "https://sports-cdn.ggops.com/social-posts-images/5ffbd0bed8b1db279dbf8c257b28a4eb.jpg", "h": 300, "media_summary": [{"regions": [{"contour": [{"y": 0.744556, "x": 0.078593}, {"y": 0.668499, "x": 0.764749}, {"y": 0.287995, "x": 0.627788}, {"y": 0.207455, "x": 0.175434}], "sub_regions": null, "features": null, "props": [{"relationships": null, "confidence": 0.5766, "confidence_min": 0.4852, "ver": "3.0", "company": "gumgum", "value": "Dog", "server": "TeamClassifier", "footprint_id": "TeamClassifier_20170203125137", "fraction": 1, "module_id": 0, "property_type": "rightsholder_name", "value_verbose": "", "property_id": 0}]}], "t2": 0, "t1": 0}], "tracks_summary": null, "w": 8, "frames_annotation": [{"regions": [{"contour": [{"y": 0.639081, "x": 0.139355}, {"y": 0.642554, "x": 0.029847}, {"y": 0.502516, "x": 0.158739}, {"y": 0.670725, "x": 0.527886}], "sub_regions": null, "features": null, "props": [{"relationships": null, "confidence": 0.85, "confidence_min": 0.7792, "ver": "3.0", "company": "gumgum", "value": "Cat", "server": "TeamClassifier", "footprint_id": "TeamClassifier_20170203125137", "fraction": 1, "module_id": 0, "property_type": "rightsholder_name", "value_verbose": "", "property_id": 0}]}], "t": 0}]}
> Context
{
"codes": [
{
"code": "SUCCESS",
"ver": "sports-production",
"company": "gumgum-sports",
"labels": [
"Cat",
"Dog"
],
"annotator": "unknown",
"server": "HAM",
"date": "20170626041248",
"tstamps": [],
"id": "HAM_20170626041248"
}
],
"url_original": "https://www.facebook.com/NYKnicks/photos/a.71815779615.98182.21410634615/10157985435429616/",
"url": "https://sports-cdn.ggops.com/social-posts-images/5ffbd0bed8b1db279dbf8c257b28a4eb.jpg",
"h": 400,
"media_summary": [
{
"regions": [
{
"contour": [
{
"y": 0.502205,
"x": 0.338184
},
{
"y": 0.503508,
"x": 0.714891
},
{
"y": 0.607232,
"x": 0.346239
},
{
"y": 0.109788,
"x": 0.467408
}
],
"sub_regions": null,
"features": null,
"props": [
{
"relationships": null,
"confidence": 1,
"confidence_min": 0,
"ver": "sports-production",
"company": "gumgum",
"value": "Cat",
"server": "HAM",
"footprint_id": "HAM_20170626041248",
"fraction": 1,
"module_id": 0,
"property_type": "rightsholder_name",
"value_verbose": "",
"property_id": 0
}
]
}
],
"t2": 0,
"t1": 0
}
],
"tracks_summary": null,
"w": 88,
"frames_annotation": [
{
"regions": [
{
"contour": [
{
"y": 0.241438,
"x": 0.889341
},
{
"y": 0.598113,
"x": 0.217485
},
{
"y": 0.948145,
"x": 0.513462
},
{
"y": 0.020427,
"x": 0.908873
}
],
"sub_regions": null,
"features": null,
"props": [
{
"relationships": null,
"confidence": 1,
"confidence_min": 0,
"ver": "sports-production",
"company": "gumgum",
"value": "Cat",
"server": "HAM",
"footprint_id": "HAM_20170626041248",
"fraction": 1,
"module_id": 0,
"property_type": "rightsholder_name",
"value_verbose": "",
"property_id": 0
}
]
}
],
"t": 0
}
]
}- Complex nested Json logs
- Ability to run complex queries
- Important Data volume:
- ~110G / day (now)
- ~200G / day (in one year)
- Long retention period (90 days)
> Experiments with Elasticsearch
Filebeat
forwarders
Elasticsearch
cluster
Analysis
interfaces
> Conclusion with Elasticsearch
- Straight forward data ingestion using filebeat
- Control of the Mapping for data structure changes and control of dynamic fields
- Possible use of Tokenizers and Analyzers
- Really good for data exploration thanks to the design of the engine
GET filebeat-*/_search
{
"query": {
"nested": {
"path": "frames_annotation",
"query": {
"nested": {
"path": "frames_annotation.regions",
"query": {
"nested": {
"path": "frames_annotation.regions.props",
"query": {
"bool": {
"must": [
{
"match": {
"frames_annotation.regions.props.property_type": "rig"
}
}
]
}
}
}
}
}
}
}
}
}> Experiments with AWS Athena and Glue
> CONCLUSION WITH AWS Athena and glue
- Need to use Glue and run ETL process in order to Relationalize the data
- Strong Data typing decreasing flexibility (confidence translates into two columns double and int)
- SQL denormalisation makes it harder to retrieve objects (one per line instead of having a single object returned)
SELECT *
FROM root, codes
WHERE root.codes = codes.id
AND codes.server = 'nbalogosdetector'
AND codes.ver = '3.1'Writing to S3 bucket: root_codes.val.labels
+---+-----+--------------------+
| id|index|codes.val.labels.val|
+---+-----+--------------------+
| 1| 0| Cat|
| 1| 1| Dog|
| 2| 0| NewYorkKnicks|
| 2| 1| ApparelSponsor|
| 2| 2| Squarespace|
| 3| 0| Cat|
| 3| 1| Dog|
| 4| null| |
| 5| null| |
| 6| null| |
| 14| 0| GoldenStateWarriors|
| 14| 1| JerseyPatch|
| 14| 2| Squarespace|
+---+-----+--------------------+
> conclusion
| AWS Athena | Elasticsearch |
|---|---|
| Not appropriate in that case | Better fits our needs |
| Probably "cheeper" | Probably "more" expensive |
| Self managed | Self managed or In-House |
> What's next...
Elasticsearch 6.x With Spotinst Elastigroup
- Reading about running ES @scale
- Reviewing ES Automations for 6.X
- Benchmark to define important cluster settings and cluster size
- Testing High availability of the cluster with Spotinst