Data retention policy

76 views
Skip to first unread message

REMI DRUILHE

unread,
Mar 24, 2020, 8:57:05 AM3/24/20
to Prometheus Users
Hello,

I am using a Docker version of Prometheus (the latest one) in which I am setting the value for the retention time (option storage.tsdb.retention.time) at launch. I trying first with small values like 1m and 60s (for 1 minute) or 5m (for 5 minutes) but I was not able to see any deletion of the data after this period. Thus, I tested again this morning with an higher value: 2h (for 2 hours). I have seen this information here (in the retention section) that 2h is the minimum value for Prometheus, not sure if it is true or not. But even with this value, the data stored 4 hours ago can still be retrieved using the HTTP API. Here is my query: curl 'http://172.18.0.3:9090/api/v1/query_range?query=go_memstats_alloc_bytes&start=2020-03-24T00:01:00.000Z&end=2020-03-24T17:00:00.000Z&step=15s'.

On the Prometheus GUI, the flag for this option is correctly setup.

Here is the Docker compose file that is used to launch Prometheus:

  prometheus:
    command: '--config.file=/etc/prometheus.yml --storage.tsdb.retention.time=2h'
    container_name: remi_prometheus
    depends_on:
    - cadvisor
    expose:
    - '9090'
    image: prom/prometheus:latest
    labels:
      project.run.user: remi
    networks:
      project-bridge:
        aliases:
        - prometheus
    ports:
    - published: 9090
      target: 9090
    volumes:
    - remi-prometheus:/prometheus:rw
    - /home/remi/Workspace/project/runtime/configuration/prometheus.yml:/etc/prometheus.yml:rw

Note that I am using cAdvisor to populate Prometheus in which I get some Golang metrics.

Thus here are my questions:
  • What am I doing wrong?
  • Was is the minimum value for storage.tsdb.retention.time?
  • Is there another option that overwrite the value set in storage.tsdb.retention.time that I am not aware of?
  • Is there a way to test that the option is working if the minimal value is 1 day? I guess it would be to change the system date, but not sure about it.

Thanks for the help,

Best regards and good luck if you are under confinement because of this damn virus :)

REMI DRUILHE

unread,
Mar 25, 2020, 5:27:22 AM3/25/20
to Prometheus Users
Anyone to help me on this subject?

Thanks.

REMI DRUILHE

unread,
Mar 26, 2020, 10:33:27 AM3/26/20
to Prometheus Users
I did some others tests, on others operating systems (not sure it is relevant because of Docker but without any clues it is better than nothing). For anyone to test it, I also changed and reduced my docker-compose file to:

version: '3.4'
services:
  prometheus:
    image: prom/prometheus:latest
    container_name: "prometheus"
    depends_on:
      - cadvisor
    command:
      - '--config.file=/etc/prometheus/prometheus.yml'
      - '--storage.tsdb.retention.time=60s'
    volumes:
      - /tmp/prometheus:/etc/prometheus
    ports:
      - "9090:9090"
    network_mode: "host"
  cadvisor:
    container_name: cadvisor
    volumes:
      - /:/rootfs:ro
      - /var/run:/var/run:ro
      - /sys:/sys:ro
      - /var/lib/docker/:/var/lib/docker:ro
      - /dev/disk/:/dev/disk:ro
    ports:
      - "8080:8080"
    network_mode: "host"

And here is my prometheus.yml

scrape_configs:
- job_name: cadvisor
  scrape_interval: 5s
  static_configs:
  - targets:
    - localhost:8080

But, the result is still the same with the following curl request: 


I can still get data that should have been deleted due to out of date of the timestamp and defined in the storage.tsdb.retention.time parameter.

According to me, it is a bug. But before opening an issue on Prometehus, I prefer to ask for advices about it.

Thanks,

Best regards.

Christian Hoffmann

unread,
Mar 26, 2020, 4:40:35 PM3/26/20
to REMI DRUILHE, Prometheus Users
Hi,

not sure if it's this what you are hitting, but the following may help:

The retention time defines how long Prometheus is expected to keep
history. This does not imply that data older than this interval will
vanish immediately.
In other words: Once the interval has passed, Prometheus is free to
delete this data, but it isn't forced to. In practice, this will depend
on maintenance (compaction) running.

This document provides further details, especially:

https://prometheus.io/docs/prometheus/latest/storage/
"Expired block cleanup happens on a background schedule. It may take up
to two hours to remove expired blocks. Expired blocks must be fully
expired before they are cleaned up."

Maybe this explains things for you?

Kind regards,
Christian
> *storage.tsdb.retention.time*) at launch. I trying first with
> small values like *1m* and *60s* (for 1 minute) or *5m* (for 5
> minutes) but I was not able to see any deletion of the data
> after this period. Thus, I tested again this morning with an
> higher value: *2h* (for 2 hours). I have seen this information
> here
> <https://medium.com/@valyala/prometheus-storage-technical-terms-for-humans-4ab4de6c3d48>
> (in the retention section) that 2h is the minimum value for
> Prometheus, not sure if it is true or not. But even with this
> value, the data stored 4 hours ago can still be retrieved using
> the HTTP API. Here is my query: *curl
> 'http://172.18.0.3:9090/api/v1/query_range?query=go_memstats_alloc_bytes&start=2020-03-24T00:01:00.000Z&end=2020-03-24T17:00:00.000Z&step=15s
> <http://172.18.0.3:9090/api/v1/query_range?query=go_memstats_alloc_bytes&start=2020-03-24T00:01:00.000Z&end=2020-03-24T17:00:00.000Z&step=15s>'*.
>
> On the Prometheus GUI, the flag for this option is correctly setup.
>
> Here is the Docker compose file that is used to launch Prometheus:
>
> |
>   prometheus:
>     command:'--config.file=/etc/prometheus.yml
> --storage.tsdb.retention.time=2h'
>     container_name:remi_prometheus
>     depends_on:
>     -cadvisor
>     expose:
>     -'9090'
>     image:prom/prometheus:latest
>     labels:
>       project.run.user:remi
>     networks:
>       project-bridge:
>         aliases:
>         -prometheus
>     ports:
>     -published:9090
>       target:9090
>     volumes:
>     -remi-prometheus:/prometheus:rw
>     -
> /home/remi/Workspace/project/runtime/configuration/prometheus.yml:/etc/prometheus.yml:rw
> |
>
> Note that I am using cAdvisor to populate Prometheus in which I
> get some Golang metrics.
>
> Thus here are my questions:
>
> * What am I doing wrong?
> * Was is the minimum value for storage.tsdb.retention.time?
> * Is there another option that overwrite the value set in
> storage.tsdb.retention.time that I am not aware of?
> * Is there a way to test that the option is working if the
> minimal value is 1 day? I guess it would be to change the
> system date, but not sure about it.
>
>
> Thanks for the help,
>
> Best regards and good luck if you are under confinement because
> of this damn virus :)
>
> --
> You received this message because you are subscribed to the Google
> Groups "Prometheus Users" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to prometheus-use...@googlegroups.com
> <mailto:prometheus-use...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/prometheus-users/c3550107-62af-4120-b2d1-13926b3fff14%40googlegroups.com
> <https://groups.google.com/d/msgid/prometheus-users/c3550107-62af-4120-b2d1-13926b3fff14%40googlegroups.com?utm_medium=email&utm_source=footer>.

REMI DRUILHE

unread,
Mar 27, 2020, 9:55:59 AM3/27/20
to Prometheus Users
Hi,

Thanks for the answer.

Just to add some context to what I want to do. We are trying to make our application (Prometheus being a part of it) compliant with GDPR. The default period for our data retention is 10 days but can be set by the user to 1h minimum. If the data are not deleted after 1 hour then, it is no longer compliant with what we declared to the national data protection authority.

From your answer, I understand now that the parameter storage.tsdb.retention.time is not hard deadline for the data. Data might live after the retention time, at least 2 hours. So I launched a new test with storage.tsdb.retention.time equals to 5m. Then, I waited for 3 hours to see if the data were deleted but they are always in Prometheus. 

Is it possible to anticipate when the data will be deleted to update ours compliance with GDPR? Can we change the background schedule to have a better estimation of when the data will be deleted? Or should I use the delete method of Prometheus form an external component to be sure that the data are deleted on time?

Thanks,

Best regards.

Stuart Clark

unread,
Mar 27, 2020, 9:58:51 AM3/27/20
to promethe...@googlegroups.com
On 27/03/2020 13:55, REMI DRUILHE wrote:
Hi,

Thanks for the answer.

Just to add some context to what I want to do. We are trying to make our application (Prometheus being a part of it) compliant with GDPR. The default period for our data retention is 10 days but can be set by the user to 1h minimum. If the data are not deleted after 1 hour then, it is no longer compliant with what we declared to the national data protection authority.

From your answer, I understand now that the parameter storage.tsdb.retention.time is not hard deadline for the data. Data might live after the retention time, at least 2 hours. So I launched a new test with storage.tsdb.retention.time equals to 5m. Then, I waited for 3 hours to see if the data were deleted but they are always in Prometheus. 

Is it possible to anticipate when the data will be deleted to update ours compliance with GDPR? Can we change the background schedule to have a better estimation of when the data will be deleted? Or should I use the delete method of Prometheus form an external component to be sure that the data are deleted on time?


What data are you expecting Prometheus to hold that would be subject to GDPR?

Prometheus stores monitoring metrics so wouldn't have anything related to personal data, so isn't in scope for GDPR...


To unsubscribe from this group and stop receiving emails from it, send an email to prometheus-use...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-users/27f3c76c-f2b3-4bfe-87d7-02adbf61f8c4%40googlegroups.com.


-- 
Stuart Clark

Ben Kochie

unread,
Mar 27, 2020, 1:39:27 PM3/27/20
to Stuart Clark, Prometheus Users
Yes, this is the general guideline. User identifiable data should not be in Prometheus. Mostly because it tends to be a cardinality problem.

Thanks,

Best regards.

--
You received this message because you are subscribed to the Google Groups "Prometheus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to prometheus-use...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-users/27f3c76c-f2b3-4bfe-87d7-02adbf61f8c4%40googlegroups.com.


-- 
Stuart Clark

--
You received this message because you are subscribed to the Google Groups "Prometheus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to prometheus-use...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-users/e056715f-cf27-5e5d-743f-686c835c8207%40Jahingo.com.

REMI DRUILHE

unread,
Mar 30, 2020, 4:34:01 AM3/30/20
to Prometheus Users
In our context, Prometheus is storing system metrics and business metrics, especially the number of accesses to the methods of our API. We do not have any direct user data stored in Prometheus. But I think that someone could cross-reference the data to find something more specific about someone (it is a big if) that is why I wanted to have a hard deadline for the deletion of the data.

Le vendredi 27 mars 2020 18:39:27 UTC+1, Ben Kochie a écrit :
Yes, this is the general guideline. User identifiable data should not be in Prometheus. Mostly because it tends to be a cardinality problem.

Thanks,

Best regards.

--
You received this message because you are subscribed to the Google Groups "Prometheus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to promethe...@googlegroups.com.


-- 
Stuart Clark

--
You received this message because you are subscribed to the Google Groups "Prometheus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to promethe...@googlegroups.com.

Brian Candler

unread,
Mar 30, 2020, 10:37:11 AM3/30/20
to Prometheus Users
On Monday, 30 March 2020 09:34:01 UTC+1, REMI DRUILHE wrote:
In our context, Prometheus is storing system metrics and business metrics, especially the number of accesses to the methods of our API.


That presumably is an aggegate of all calls to a particular method.

If you recorded counts as separate metrics labelled by source IP address or username, then that would be identifiable.  But prometheus does not work well with such high cardinality metrics anyway.

REMI DRUILHE

unread,
Mar 31, 2020, 7:27:39 AM3/31/20
to Prometheus Users
Yeah, it is just the timestamp of the call that is stored, not the IP or the user name. Thus, it is not identifiable with Prometheus only. But, the system aims at being used by 1 or 2 persons at the same time in a closed network. In this context, I think it could be easy for someone to associate the timestamp with the person that was using the application at a specific time.

Anyway, I will figure out another way to achieve what we would like to do.

Thanks for the help.

Stuart Clark

unread,
Mar 31, 2020, 7:40:03 AM3/31/20
to REMI DRUILHE, Prometheus Users
How are you storing the timestamp? Is that in a label or a metric value as the last call to the API?

In general these are sounding like you are trying to store events within Prometheus rather than metrics. Normally you'd not have a timestamp but a counter of the number of calls to the API.
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

REMI DRUILHE

unread,
Mar 31, 2020, 10:27:36 AM3/31/20
to Prometheus Users
In our code, we are using a counter to count the accesses to the various methods of the API. We have one counter per method. We do not store the timestamp. But when we ask Prometheus with a "query_range" (see request below), it returns the list of all the methods that have been accessed. 

For each of our API method, it also returns a list of key-value where the key is the timestamp and the value is the value of the counter at that time (see example below). Thus, in some way, you are able to track when the method has been called. And if our system is used by a single user, then it is easy to follow which methods he called. It is a bit twisted, but the national data protection authority might also be twisted sometimes... But according to your previous answers, maybe we did not used the counter in a proper way and we should change the way it is designed.

{
   "status":"success",
   "data":{
      "resultType":"matrix",
      "result":[
         {
            "metric":{
               "__name__":"bea_nb_request",
               "action":"my_api_method",
               "instance":"bea:8081",
               "job":"bea"
            },
            "values":[
               [
                  1585663440,
                  "1"
               ],
               [
                  1585663500,
                  "2"
               ],
               [
                  1585663560,
                  "3"
               ],
               [
                  1585663620,
                  "3"
               ],
               [
                  1585663680,
                  "3"
               ],
               [
                  1585663740,
                  "3"
               ],
               [
                  1585663800,
                  "3"
               ],
               [
                  1585663860,
                  "3"
               ]
            ]
         },
         others_api_methods...
         }
      ]
   }
}



Le mardi 31 mars 2020 13:40:03 UTC+2, Stuart Clark a écrit :
How are you storing the timestamp? Is that in a label or a metric value as the last call to the API?

In general these are sounding like you are trying to store events within Prometheus rather than metrics. Normally you'd not have a timestamp but a counter of the number of calls to the API.

Stuart Clark

unread,
Mar 31, 2020, 11:51:31 AM3/31/20
to REMI DRUILHE, Prometheus Users
No that sounds fairly normal. One thing to note is that those timestamps are not the times the methods were called. They are when Prometheus scraped your application. So if you scrape once a minute the actual call could have been at any point during that minute. Equally if there are multiple calls during that minute you'd have no idea when they happened either.

I'm not a lawyer or GDPR expert, but I think the type of extreme de-anonymisation you are suggesting is not generally something you'd be expected to be worrying about. Equally even if you do have an idea of who might have called an API there still isn't any personal data in Prometheus.

REMI DRUILHE

unread,
Apr 1, 2020, 4:58:55 AM4/1/20
to Prometheus Users
You are right, most of the time this kind de-anonymisation is extreme. And right again when you say that there is no personal data stored in Prometheus.

I am also not a lawyer but I know from my DPO that the national data protection authority (NDPA) might be very very very meticulous, especially in my domain of video processing... We had several meetings about it and we had to review our data processing multiple time. I was just looking for a way to delete data with a hard deadline if the NDPA say that the current solution is not good enough (the one with storage.tsdb.retention.time option). I think it is better to come with an answer than saying that we did not thought about it.


Le mardi 31 mars 2020 17:51:31 UTC+2, Stuart Clark a écrit :
No that sounds fairly normal. One thing to note is that those timestamps are not the times the methods were called. They are when Prometheus scraped your application. So if you scrape once a minute the actual call could have been at any point during that minute. Equally if there are multiple calls during that minute you'd have no idea when they happened either.

I'm not a lawyer or GDPR expert, but I think the type of extreme de-anonymisation you are suggesting is not generally something you'd be expected to be worrying about. Equally even if you do have an idea of who might have called an API there still isn't any personal data in Prometheus.

Stuart Clark

unread,
Apr 1, 2020, 5:12:01 AM4/1/20
to REMI DRUILHE, Prometheus Users
On 2020-04-01 09:58, REMI DRUILHE wrote:
> You are right, most of the time this kind de-anonymisation is extreme.
> And right again when you say that there is no personal data stored in
> Prometheus.
>
> I am also not a lawyer but I know from my DPO that the national data
> protection authority (NDPA) might be very very very meticulous,
> especially in my domain of video processing... We had several meetings
> about it and we had to review our data processing multiple time. I was
> just looking for a way to delete data with a hard deadline if the NDPA
> say that the current solution is not good enough (the one with
> storage.tsdb.retention.time option). I think it is better to come with
> an answer than saying that we did not thought about it.

Unfortunately there are no guarantees around deletion.

In addition to the fuzziness around exactly when a block might be
removed you can also end up with data files hanging around in certain
error scenarios (e.g. tmp files if there are issues loading the WAL on
startup or during block rotation)

>
> Le mardi 31 mars 2020 17:51:31 UTC+2, Stuart Clark a écrit :
>
>> No that sounds fairly normal. One thing to note is that those
>> timestamps are not the times the methods were called. They are when
>> Prometheus scraped your application. So if you scrape once a minute
>> the actual call could have been at any point during that minute.
>> Equally if there are multiple calls during that minute you'd have no
>> idea when they happened either.
>>
>> I'm not a lawyer or GDPR expert, but I think the type of extreme
>> de-anonymisation you are suggesting is not generally something you'd
>> be expected to be worrying about. Equally even if you do have an
>> idea of who might have called an API there still isn't any personal
>> data in Prometheus.
>>
>> On 31 March 2020 15:27:36 BST, REMI DRUILHE <remi....@atos.net>
>> wrote:
>> In our code, we are using a counter to count the accesses to the
>> various methods of the API. We have one counter per method. We do
>> not store the timestamp. But when we ask Prometheus with a
>> "query_range" (see request below), it returns the list of all the
>> methods that have been accessed.
>>
>> curl
>>
> 'http://172.22.0.15:9090/api/v1/query_range?query=bea_nb_request&start=2020-03-31T00:01:00.000Z&end=2020-03-31T17:00:00.000Z&step=60s
>> [1]'
> --
> You received this message because you are subscribed to the Google
> Groups "Prometheus Users" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to prometheus-use...@googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/prometheus-users/29f1e59c-1d72-436a-9883-c81c71e0cbd9%40googlegroups.com
> [2].
>
>
> Links:
> ------
> [1]
> http://172.22.0.15:9090/api/v1/query_range?query=bea_nb_request&amp;start=2020-03-31T00:01:00.000Z&amp;end=2020-03-31T17:00:00.000Z&amp;step=60s
> [2]
> https://groups.google.com/d/msgid/prometheus-users/29f1e59c-1d72-436a-9883-c81c71e0cbd9%40googlegroups.com?utm_medium=email&utm_source=footer

--
Stuart Clark

REMI DRUILHE

unread,
Apr 1, 2020, 9:30:58 AM4/1/20
to Prometheus Users
No problem :)

At least it is pretty clear for me now. I will discuss with my DPO again to have its point of view on this. Maybe I am over-reacting about those data.

Anyway, thanks for the help and the talk. It has been very interesting.
Reply all
Reply to author
Forward
0 new messages