@cbrassel please provide "info memory" output from 1.17.1 and then 1.18.0 after 24 hours of running both versions

from dragonfly.

This is the 1.17.1 memory output :

Memory

used_memory:4876027424
used_memory_human:4.54GiB
used_memory_peak:4876027424
used_memory_peak_human:4.54GiB
fibers_stack_vms:1604848
fibers_count:49
used_memory_rss:5124042752
used_memory_rss_human:4.77GiB
used_memory_peak_rss:5183430656
maxmemory:54166546022
maxmemory_human:50.45GiB
object_used_memory:4411032760
type_used_memory_STRING:4393906584
type_used_memory_ZSET:17126176
table_used_memory:453446176
num_buckets:807780
num_entries:7958770
inline_keys:597957
listpack_blobs:0
listpack_bytes:0
small_string_bytes:3436687384
pipeline_cache_bytes:8242
dispatch_queue_bytes:0
dispatch_queue_subscriber_bytes:0
dispatch_queue_peak_bytes:26415220
client_read_buffer_peak_bytes:1728512
tls_bytes:5664
cache_mode:store
maxmemory_policy:noeviction

As mentioned it is a production system, so knowing that they can be a production outage I can not just put the new release in production and hoping .. sorry for that

from dragonfly.

np, we have not seen anything irregular with Dragonfly's memory consumption in 1.18. we will keep looking.
In any case, this issue unfortunately does not have any actionable inputs, so we do not know what to fix.
Closing for now. Feel free to reopen if you succeed to get more info.

from dragonfly.

Actually, we reproduced the problem.

from dragonfly.

@cbrassel we confirm there is a problem with socket leakage in kernel. The leakage is caused due to the new io_uiring component that was enabled by default in 1.18. Luckily it's easy to fix by passing --enable_direct_fd=false flag when running dragonfly. We will also release 1.18.1 with this fix soon.

from dragonfly.