Prometheus’s simple and reliable operational model is one of its major selling points. However, after surpassing a certain scale, we have identified a few shortcomings it imposes. We are proud to present Thanos, an open source project by Improbable that bundles a set of components that seamlessly transform existing Prometheus deployments, into a unified, global scale monitoring system. Authors: Fabian Reinartz, Bartlomiej Plotka Slides from January London Prometheus Meetup 2018. Thanos: https://github.com/improbable-eng/thanos

1. Bartek Plotka
Bwplotka Bplotka
Fabian Reinartz
fabxc
Global, durable Prometheus monitoring

2. Prometheus 2.0
● Reliable operational model
● Powerful query language
● Scraping capabilities beyond the casual usage
● Local metric storage

3. Prometheus at Scale
A dream

4. n1
Improbable case
2 n+1
...
● Multiple isolated Kubernetes clusters

5. n1
Improbable case
2
Prometheus
n+1
Prometheus
...
● Multiple isolated Kubernetes clusters
● Single Prometheus server per cluster

6. 1
Improbable case
2
Prometheus
n
n+1
Prometheus
...
● Multiple isolated Kubernetes clusters
● Single Prometheus server per cluster
● Dashboards & Alertmanager in separate cluster
Grafana
Alertmanager

7. Improbable case
● Multiple isolated Kubernetes clusters
● Single Prometheus server per cluster
● Dashboards & Alertmanager in separate cluster
Grafana
Alertmanager
What is missing? 1
2
Prometheus
n
n+1
Prometheus
...

8. Global View
See everything from a single
place!

9. 1
Global View
2
Prometheus
n
n+1
Prometheus
...
Grafana
Alertmanager
“Alert when 66% of clusters in a region are down”

10. 1
Global View
2
Prometheus
n
n+1
Prometheus
...
● How to aggregate data from different clusters?
Grafana
Alertmanager

11. 1
Global View
2
Prometheus
n
n+1
Prometheus
...
● How to aggregate data from different clusters?
○ Use hierarchical federation?
Grafana
Alertmanager
Prometheus

12. 1
Global View
2
Prometheus
n
n+1
Prometheus
...
● How to aggregate data from different clusters?
○ Use hierarchical federation?
Grafana
AlertmanagerSingle point of failure
Maintenance
What data are federated?
Prometheus

13. Availability
Where is my sample?

14. 1
Availability
2
Prometheus
n
n+1
Prometheus
...
Grafana
Alertmanager
Operator error
Hardware failure
Rollout

15. 1
Availability
2
Prometheus
n
n+1
Prometheus
...
● Can we assure no loss in metric data?
Grafana
Alertmanager

16. 1
Availability
2
Prometheus
n
n+1
Prometheus
...
● Can we assure no loss in metric data?
○ Add HA replicas?
Grafana
Alertmanager
Prometheus Prometheus

17. 1
Availability
2
Prometheus
n
n+1
Prometheus
...
● Can we assure no loss in metric data?
○ Add HA replicas?
Grafana
Alertmanager
Prometheus Prometheus
What replica we should query?
Cost?
Where to put rules and alerts?

18. Historical Metrics
What exactly happened
X weeks ago?

19. Metric retention
T - 3 years TT - 12 monthsT - 2 years
?
“Go back to what happened 6 months ago...”

20. Metric retention
T - 3 years TT - 12 monthsT - 2 years
?
“Good progress! Memory allocs looks better than 1,5 year ago!”
T
X

21. Metric retention
“Let’s see user traffic across years!”
T - 2 years T - 12 monthsT - 3 years
? ? ? ?
T
X

22. Metric retention
Infrastructure retention: 9 days
9 days
T - 3 years TT - 12 monthsT - 2 years T

23. Metric retention
● Can we have longer retention?

24. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0

25. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0
○ Scale SSD Vertically?
SSD
Prometheus

26. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0
○ Scale SSD Vertically?
SSD
Prometheus

27. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0
○ Scale SSD Vertically?
SSD
Prometheus

28. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0
○ Scale SSD Vertically?
SSD
Prometheus
SSD
SSD
SSD
SSD

29. Metric retention
● Can we have longer retention?
○ Upgrade to Prometheus 2.0
○ Scale SSD Vertically?
SSD
Prometheus
Backup?
Maintenance
Cost

30. Recap
1
2
Prometheus
n
n+1
Prometheus
...
Grafana
Alertmanager
Prometheus
It is just hard to…
● Have a global view
● Have a HA in place
● Increase retention

31. Thanos
It is just hard to…
● Have a global view
● Have a HA in place
● Increase retention

32. ● Seamless integration with Prometheus
● Easy deployment model
● Minimal number of dependencies
● Minimal baseline cost
Additional Goals

33. Global View
See everything from a single
place!

34. SSD
Prometheus
Prometheus
Targets

35. SSD
Sidecar
Prometheus Sidecar
Targets

36. SSD
Sidecar
Prometheus Sidecar
Targets
gRPC (Store API)

37. Store API
service Store {
rpc Series(SeriesRequest) returns (stream SeriesResponse);
rpc LabelNames(LabelNamesRequest) returns (LabelNamesResponse);
rpc LabelValues(LabelValuesRequest) returns (LabelValuesResponse);
}
message SeriesRequest {
int64 min_time = 1;
int64 max_time = 2;
repeated LabelMatcher matchers = 3;
}
Sidecar
Prometheus
remote read
Store API

38. SSD
Querier
Prometheus Sidecar
Querier
Store API
Targets
HTTP
Query API

39. SSD
Global View
Prometheus Sidecar
Querier
Targets
SSD
Sidecar
Targets
Prometheus
Merge
Store API

40. SSD
Global View + Availability
Prometheus Sidecar
Targets
SSD
Sidecar
Targets
Prometheus
SSD
Sidecar Prometheus
“replica”:”1”
“replica”:”2”
Querier
Merge
Deduplicate
Store API

41. Thanos
It is just hard to…
● Have a global view
● Have a HA in place
● Increase retention

42. Historical Metrics
What exactly happened
X weeks ago?

43. TSDB Layout
Block 2 Block 4Block 3Block 1
T-200T-300 T-100 T-50 T

44. TSDB Layout
Block 4Block 3Block 1
chunks chunks
chunks chunks
index
T-200T-300 T-100 T-50 T

45. SSD
Data saving
Prometheus Sidecar
Targets
Object Storage
Blocks Blocks
Block

46. Store
Object Storage
Blocks
Cache
Store
Querier
Store API

47. Store
Object Storage
Blocks
Cache
Store
Querier
Block
Store API

48. Store
● A series is made up of one or more “chunks”
● A chunk contains ~120 samples each
● Chunks can be retrieved through HTTP byte
range queries

49. Store
● A series is made up of one or more “chunks”
● A chunk contains ~120 samples each
● Chunks can be retrieved through HTTP byte
range queries
Example:
● 1000 series @ 30s scrape interval

50. Store
● A series is made up of one or more “Chunks”
● A chunk contains ~120 samples each
● Chunks can be retrieved through HTTP byte
range queries
Example:
● 1000 series @ 30s scrape interval
● Query 1 year
8.7 million chunks/range queries

51. Store
Leverage Prometheus’ TSDB file layout

52. Store
Leverage Prometheus’ TSDB file layout
● Chunks of the same series are aligned
sequentially

53. Store
Leverage Prometheus’ TSDB file layout
● Chunks of the same series are aligned
● Similar series are aligned, e.g. due to same
metric name

54. Store
Leverage Prometheus’ TSDB file layout
● Chunks of the same series are aligned
● Similar series are aligned, e.g. due to same
metric name
Consolidating ranges in close proximity reduces
request count by 4-6 orders of magnitude.
8.7 million requests turned into O(20) requests.

55. Store
Leverage Prometheus’ TSDB file layout
● Chunks of the same series are aligned
● Similar series are aligned, e.g. due to same
metric name
Index lookups profit from a similar approach.

56. Compaction
Density matters

57. Compaction
Object Storage
Blocks
Disk
Compactor

58. Compaction
Object Storage
Blocks
Disk
Compactor
Blocks

59. Compaction
Object Storage
Blocks
Disk
Compactor
Blocks
Block

60. Compaction
Object Storage
Blocks
Disk
Compactor
Block

61. Thanos
It is just hard to…
● Have a global view
● Have a HA in place
● Increase retention

62. Downsampling
Let’s just step back a little

63. Downsampling
Raw: 16 bytes/sample
Compressed: 1.07
bytes/sample

64. Downsampling
BUT…

65. Downsampling
Decompressing one sample takes 10-40 nanoseconds
● Times 1000 series @ 30s scrape interval
● Times 1 year

66. Downsampling
Decompressing one sample takes 10-40 nanoseconds
● Times 1000 series @ 30s scrape interval
● Times 1 year
● Over 1 billion samples, i.e. 10-40s – for decoding alone
● Plus your actual computation over all those samples, e.g. rate()

67. Downsampling
Block
RAW
Block
@ 5m
Block
@ 1h
10x 12x

68. Downsampling
raw chunk
count sum min max counter
raw chunk...

69. Downsampling
count sum min max counter
...

70. Downsampling
count sum min max counter
count_over_time(requests_total[1h])

71. Downsampling
count sum min max counter
sum_over_time(requests_total[1h])

72. Downsampling
count sum min max counter
min(requests_total)
min_over_time(requests_total[1h])

73. Downsampling
count sum min max counter
max(requests_total)
max_over_time(requests_total[1h])

74. Downsampling
count sum min max counter
rate(requests_total[1h])
increase(requests_total[1h])

75. Downsampling
count sum min max counter
requests_total
avg(requests_total)
...
*
avg

76. Full Architecture
Querier
SSD
Sidecar Prometheus
SSD
Sidecar Prometheus
QuerierQuerier
…
Compactor
Store
Bucket

77. Full Architecture
$ thanos sidecar …
$ thanos query …
$ thanos store …
$ thanos compact …

78. Deployment Models
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Cluster A
Cluster B
Cluster C

79. Deployment Models
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
Cluster A
Cluster B
Cluster C
Federation (through Store API)

80. Deployment Models
Querier
S P
QuerierQuerier
…
Store
Bucket
S P
S P …
Store
Bucket
S P
S P …
Store
Bucket
S P
Cluster A
Cluster B
Cluster C
Global Scale Thanos Cluster

81. Cost
● Store + Query node ~ Savings on Prometheus side (+/- 0)
● Fewer SSD space on Prometheus side (savings)
● Basically: just your data stored in S3/GCS/HDFS + requests

82. Cost
Example:
● 20 Prometheus servers each ingesting 100k samples/sec, 500GB of local disk
● 20 x 250GB of new data per month + ~20% overhead for downsampling
● $1440/month for storage after 1 year (72TB of queryable data)
● $100/month for sustained 100 query/sec against object storage
Thanos Cost: $1540

83. Cost
Example:
● 20 Prometheus servers each ingesting 100k samples/sec, 500GB of local disk
● 20 x 250GB of new data per month + ~20% overhead for downsampling
● $1440/month for storage after 1 year (72TB of queryable data)
● $100/month for sustained 100 query/sec against object storage
● $1530/month savings in local SSDs
Thanos Cost: $1540 Prometheus Savings: $1530

84. Demo - retention

85. Demo - deduplication

86. Demo - deduplication

87. Any questions?
github.com/improbable-eng/thanos
Fabian Reinartz
fabxc
Bartek Plotka
bwplotka Bplotka