This document describes known failure modes of promtail
on edge cases and the
adopted trade-offs.
Given the following order of events:
promtail
is tailing/app.log
promtail
current position for/app.log
is100
(byte offset)promtail
is stopped/app.log
is truncated and new logs are appended to itpromtail
is restarted
When promtail
is restarted, it reads the previous position (100
) from the
positions file. Two scenarios are then possible:
/app.log
size is less than the position before truncating/app.log
size is greater than or equal to the position before truncating
If the /app.log
file size is less than the previous position, then the file is
detected as truncated and logs will be tailed starting from position 0
.
Otherwise, if the /app.log
file size is greater than or equal to the previous
position, promtail
can't detect it was truncated while not running and will
continue tailing the file from position 100
.
Generally speaking, promtail
uses only the path to the file as key in the
positions file. Whenever promtail
is started, for each file path referenced in
the positions file, promtail
will read the file from the beginning if the file
size is less than the offset stored in the position file, otherwise it will
continue from the offset, regardless the file has been truncated or rolled
multiple times while promtail
was not running.
For each tailing file, promtail
reads a line, process it through the
configured pipeline_stages
and push the log entry to Loki. Log entries are
batched together before getting pushed to Loki, based on the max batch duration
client.batch-wait
and size client.batch-size-bytes
, whichever comes first.
In case of any error while sending a log entries batch, promtail
adopts a
"retry then discard" strategy:
promtail
retries to send log entry to the ingester up tomaxretries
times- If all retries fail,
promtail
discards the batch of log entries (which will be lost) and proceeds with the next one
You can configure the maxretries
and the delay between two retries via the
backoff_config
in the promtail config file:
clients:
- url: INGESTER-URL
backoff_config:
minbackoff: 100ms
maxbackoff: 10s
maxretries: 10
The following table shows an example of the total delay applied by the backoff algorithm
with minbackoff: 100ms
and maxbackoff: 10s
:
Retry | Min delay | Max delay | Total min delay | Total max delay |
---|---|---|---|---|
1 | 100ms | 200ms | 100ms | 200ms |
2 | 200ms | 400ms | 300ms | 600ms |
3 | 400ms | 800ms | 700ms | 1.4s |
4 | 800ms | 1.6s | 1.5s | 3s |
5 | 1.6s | 3.2s | 3.1s | 6.2s |
6 | 3.2s | 6.4s | 6.3s | 12.6s |
7 | 6.4s | 10s | 12.7s | 22.6s |
8 | 6.4s | 10s | 19.1s | 32.6s |
9 | 6.4s | 10s | 25.5s | 42.6s |
10 | 6.4s | 10s | 31.9s | 52.6s |
11 | 6.4s | 10s | 38.3s | 62.6s |
12 | 6.4s | 10s | 44.7s | 72.6s |
13 | 6.4s | 10s | 51.1s | 82.6s |
14 | 6.4s | 10s | 57.5s | 92.6s |
15 | 6.4s | 10s | 63.9s | 102.6s |
16 | 6.4s | 10s | 70.3s | 112.6s |
17 | 6.4s | 10s | 76.7s | 122.6s |
18 | 6.4s | 10s | 83.1s | 132.6s |
19 | 6.4s | 10s | 89.5s | 142.6s |
20 | 6.4s | 10s | 95.9s | 152.6s |
When promtail
shuts down gracefully, it saves the last read offsets in the
positions file, so that on a subsequent restart it will continue tailing logs
without duplicates neither losses.
In the event of a crash or abruptly termination, promtail
can't save the last
read offsets in the positions file. When restarted, promtail
will read the
positions file saved at the last sync period and will continue tailing the files
from there. This means that if new log entries have been read and pushed to the
ingester between the last sync period and the crash, these log entries will be
sent again to the ingester on promtail
restart.
However, for each log stream (set of unique labels) the Loki ingester skips all
log entries received out of timestamp order. For this reason, even if duplicated
logs may be sent from promtail
to the ingester, entries whose timestamp is
older than the latest received will be discarded to avoid having duplicated
logs. To leverage this, it's important that your pipeline_stages
include
the timestamp
stage, parsing the log entry timestamp from the log line instead
of relying on the default behaviour of setting the timestamp as the point in
time when the line is read by promtail
.