diff --git a/CHANGELOG.md b/CHANGELOG.md
index e473826b8..acc668c91 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -85,6 +85,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 
+* **role:postgresql_server**: The daily dump runs with low CPU and IO priority, so it no longer slows down PostgreSQL.
+* **role:mongodb**: The daily dump runs with low CPU and IO priority, so it no longer slows down MongoDB.
+* **role:mariadb_server**: The daily dump runs with low CPU and IO priority, so it no longer slows down MariaDB.
+* **role:influxdb**: The daily dump runs with low CPU and IO priority, so it no longer slows down InfluxDB.
+* **role:gitlab_ce**: The daily dump runs with low CPU and IO priority, so it no longer slows down GitLab.
 * **role:redis, role:valkey**: Redis and Valkey also listen on a Unix socket that only the members of their group may use, on RHEL at the path the package ships (`/run/redis/redis.sock`, `/run/valkey/valkey.sock`).
 * **role:apache_httpd**: Apache answers requests for PHP files that do not exist with 404 itself instead of passing them to PHP-FPM, so scanners no longer tie up pool workers or fill the error log.
 * **playbook:setup_basic**: Checks the variables of all its roles, the compiled firewall in the Firewall Builder repository and the reachability of the Icinga2 master before it changes anything on the host, so a missing prerequisite no longer aborts the run halfway through, after SSH was already hardened.
@@ -125,6 +130,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+* **role:mariadb_server**: The daily dump works with the mydumper that Debian and Ubuntu ship.
 * **role:lvm**: `growpart: true` grows the right partition when the PV is partition 10 or higher, such as `/dev/vda12` or `/dev/nvme0n1p12`. Until now it addressed partition 2 there, or failed.
 * **role:lvm**: The argument specs name the `fstype` subkey correctly and accept `lvm__lvs__dependent_var` and `lvm__vgs__dependent_var`, so the `setup_*` playbooks can inject LVs and VGs.
 * **role:repo_mongodb**: On RHEL, the repository key always matches `repo_mongodb__version`, instead of reusing the key of an earlier run with another version, which made the installation fail.
diff --git a/extensions/molecule/setup_nextcloud/inventory/group_vars/systems_under_test.yml b/extensions/molecule/setup_nextcloud/inventory/group_vars/systems_under_test.yml
index aaccef6e0..5915dd03b 100644
--- a/extensions/molecule/setup_nextcloud/inventory/group_vars/systems_under_test.yml
+++ b/extensions/molecule/setup_nextcloud/inventory/group_vars/systems_under_test.yml
@@ -15,6 +15,11 @@ icinga2_master__downtime_api_user:
 mariadb_server__admin_user:
   username: 'mariadb-admin'
   password: 'linuxfabrik'
+# setting the dump user enables the daily dump (mariadb-dump.service and .timer)
+mariadb_server__dump_user:
+  username: 'mariadb-backup'
+  password: 'linuxfabrik'
+  state: 'present'
 
 nextcloud__fqdn: 'nextcloud.example.com'
 nextcloud__skip_notify_push: true
diff --git a/extensions/molecule/setup_nextcloud/verify.yml b/extensions/molecule/setup_nextcloud/verify.yml
index 7735e45cd..54ed7b941 100644
--- a/extensions/molecule/setup_nextcloud/verify.yml
+++ b/extensions/molecule/setup_nextcloud/verify.yml
@@ -73,3 +73,34 @@
         fail_msg: 'mariadbd runs as "{{ __molecule__mariadbd_label_result["stdout"] | trim }}" instead of mysqld_t, so the mysqld_exec_t label on the binary is missing.'
       when:
         - '__molecule__mariadbd_label_result is not skipped'
+
+
+# The dump must not slow down MariaDB. Ask systemd for the effective values instead of reading the
+# unit file, so a unit that fails to load or a setting it rejects shows up here.
+- name: 'Verify mariadb-dump.service runs with low CPU and IO priority'
+  hosts: 'systems_under_test'
+  gather_facts: false
+  become: true
+  tasks:
+
+    - name: 'systemctl show mariadb-dump.service'
+      ansible.builtin.command: >-
+        systemctl show mariadb-dump.service
+        --property=CPUWeight,IOSchedulingClass,IOSchedulingPriority,IOWeight,Nice
+      register: '__molecule__mariadb_dump_show_result'
+      changed_when: false
+
+    - name: 'Assert the dump runs niced, best-effort and with a low CPU and IO weight'
+      ansible.builtin.assert:
+        that:
+          - '"CPUWeight=20" in __molecule__mariadb_dump_show_result["stdout_lines"]'
+          - '"IOSchedulingClass=2" in __molecule__mariadb_dump_show_result["stdout_lines"]'  # 2 = best-effort
+          - '"IOSchedulingPriority=7" in __molecule__mariadb_dump_show_result["stdout_lines"]'
+          - '"IOWeight=20" in __molecule__mariadb_dump_show_result["stdout_lines"]'
+          - '"Nice=19" in __molecule__mariadb_dump_show_result["stdout_lines"]'
+        fail_msg: 'mariadb-dump.service has unexpected priorities: {{ __molecule__mariadb_dump_show_result["stdout_lines"] }}'
+
+    # the dump itself has to work with the mydumper of each platform, not just carry the settings
+    - name: 'systemctl start mariadb-dump.service'
+      ansible.builtin.command: 'systemctl start mariadb-dump.service'
+      changed_when: false
diff --git a/roles/gitlab_ce/README.md b/roles/gitlab_ce/README.md
index fb1d4d7c8..a2a74ed64 100644
--- a/roles/gitlab_ce/README.md
+++ b/roles/gitlab_ce/README.md
@@ -9,6 +9,11 @@ This role installs and configures [GitLab CE](https://about.gitlab.com/), includ
 *Available since LFOps `2.0.0`.*
 
 
+## How the Role Behaves
+
+* `gitlab-dump.service` runs with `Nice=19`, best-effort IO priority 7 and `CPUWeight=` / `IOWeight=` of `20` (systemd default: `100`), so the dump does not slow down the application on a busy host. The weights only take effect under contention. `IOWeight=` only has an effect on block devices that use the BFQ IO scheduler; with `none` or `mq-deadline`, as on a RHEL 9 VM with virtio disks, it is ignored.
+
+
 ## Dependent Roles
 
 Any [LFOps playbook](https://github.com/Linuxfabrik/lfops/blob/main/playbooks/README.md) that installs this role runs these for you. Optional ones can be disabled via the playbook's skip variables.
diff --git a/roles/gitlab_ce/templates/etc/systemd/system/gitlab-dump.service.j2 b/roles/gitlab_ce/templates/etc/systemd/system/gitlab-dump.service.j2
index 6cea073f0..25786b972 100644
--- a/roles/gitlab_ce/templates/etc/systemd/system/gitlab-dump.service.j2
+++ b/roles/gitlab_ce/templates/etc/systemd/system/gitlab-dump.service.j2
@@ -1,5 +1,5 @@
 # {{ ansible_managed }}
-# 2022091501
+# 2026101301
 
 [Unit]
 Description=Dumps (Backups) for GitLab
@@ -10,3 +10,10 @@ Description=Dumps (Backups) for GitLab
 ExecStart=-/usr/bin/gitlab-backup create
 Type=oneshot
 User=root
+# the dump must not slow down GitLab and the applications behind it. Nice= and the CPU / IO
+# weights (default 100) only matter under contention, so an idle host dumps at full speed.
+Nice=19
+IOSchedulingClass=best-effort
+IOSchedulingPriority=7
+CPUWeight=20
+IOWeight=20
diff --git a/roles/influxdb/README.md b/roles/influxdb/README.md
index 3a7e3b11a..8189ced8a 100644
--- a/roles/influxdb/README.md
+++ b/roles/influxdb/README.md
@@ -6,6 +6,11 @@ This role installs and configures [InfluxDB](https://www.influxdata.com/products
 *Available since LFOps `2.0.0`.*
 
 
+## How the Role Behaves
+
+* `influxdb-dump.service` runs with `Nice=19`, best-effort IO priority 7 and `CPUWeight=` / `IOWeight=` of `20` (systemd default: `100`), so the dump does not slow down the application on a busy host. The weights only take effect under contention. `IOWeight=` only has an effect on block devices that use the BFQ IO scheduler; with `none` or `mq-deadline`, as on a RHEL 9 VM with virtio disks, it is ignored.
+
+
 ## Dependent Roles
 
 Any [LFOps playbook](https://github.com/Linuxfabrik/lfops/blob/main/playbooks/README.md) that installs this role runs these for you. Optional ones can be disabled via the playbook's skip variables.
diff --git a/roles/influxdb/templates/etc/systemd/system/influxdb-dump.service.j2 b/roles/influxdb/templates/etc/systemd/system/influxdb-dump.service.j2
index fc6760e7a..149a0d423 100644
--- a/roles/influxdb/templates/etc/systemd/system/influxdb-dump.service.j2
+++ b/roles/influxdb/templates/etc/systemd/system/influxdb-dump.service.j2
@@ -1,5 +1,5 @@
 # {{ ansible_managed }}
-# 2022050901
+# 2026101301
 
 [Unit]
 Description=Dumps (Backups) for InfluxDB
@@ -8,3 +8,10 @@ Description=Dumps (Backups) for InfluxDB
 ExecStart=/usr/local/bin/influxdb-dump
 Type=oneshot
 User=root
+# the dump must not slow down InfluxDB and the applications behind it. Nice= and the CPU / IO
+# weights (default 100) only matter under contention, so an idle host dumps at full speed.
+Nice=19
+IOSchedulingClass=best-effort
+IOSchedulingPriority=7
+CPUWeight=20
+IOWeight=20
diff --git a/roles/mariadb_server/README.md b/roles/mariadb_server/README.md
index 906128726..4a7471f0e 100644
--- a/roles/mariadb_server/README.md
+++ b/roles/mariadb_server/README.md
@@ -25,6 +25,11 @@ By default, Ansible runs each task on all hosts affected by a play before starti
 *Available since LFOps `2.0.0`.*
 
 
+## How the Role Behaves
+
+* `mariadb-dump.service` runs with `Nice=19`, best-effort IO priority 7 and `CPUWeight=` / `IOWeight=` of `20` (systemd default: `100`), so the dump does not slow down the application on a busy host. The weights only take effect under contention. `IOWeight=` only has an effect on block devices that use the BFQ IO scheduler; with `none` or `mq-deadline`, as on a RHEL 9 VM with virtio disks, it is ignored. The IO class is deliberately not `idle`: `mydumper` holds a read lock on every non-InnoDB table (MyISAM, Aria, most of `mysql.*`) until it has dumped that table, and an idle dump could stretch that lock and keep writes waiting.
+
+
 ## Dependent Roles
 
 Any [LFOps playbook](https://github.com/Linuxfabrik/lfops/blob/main/playbooks/README.md) that installs this role runs these for you. Optional ones can be disabled via the playbook's skip variables.
diff --git a/roles/mariadb_server/templates/etc/systemd/system/mariadb-dump.service.j2 b/roles/mariadb_server/templates/etc/systemd/system/mariadb-dump.service.j2
index 0fccb86bb..daa261ed5 100644
--- a/roles/mariadb_server/templates/etc/systemd/system/mariadb-dump.service.j2
+++ b/roles/mariadb_server/templates/etc/systemd/system/mariadb-dump.service.j2
@@ -1,5 +1,5 @@
 # {{ ansible_managed }}
-# 2020112301
+# 2026101301
 
 [Unit]
 Description=mariadbdump Service
@@ -8,6 +8,16 @@ Description=mariadbdump Service
 ExecStart=/usr/local/bin/mariadb-dump
 Type=oneshot
 User=root
+# the dump must not slow down MariaDB and the applications behind it. Nice= and the CPU / IO
+# weights (default 100) only matter under contention, so an idle host dumps at full speed.
+# best-effort instead of idle: mydumper holds `LOCK TABLES ... READ LOCAL` on every non-InnoDB
+# table (MyISAM, Aria, most of `mysql.*`) until that table is dumped, so a starved dump keeps
+# writes to these tables waiting. verified against mydumper 1.0.11 and MariaDB 11.8 on Rocky 9.
+Nice=19
+IOSchedulingClass=best-effort
+IOSchedulingPriority=7
+CPUWeight=20
+IOWeight=20
 
 [Install]
 WantedBy=basic.target
diff --git a/roles/mariadb_server/templates/usr/local/bin/mariadb-dump.j2 b/roles/mariadb_server/templates/usr/local/bin/mariadb-dump.j2
index 8354457e8..e430c131b 100644
--- a/roles/mariadb_server/templates/usr/local/bin/mariadb-dump.j2
+++ b/roles/mariadb_server/templates/usr/local/bin/mariadb-dump.j2
@@ -6,7 +6,7 @@
 # License: The Unlicense, see LICENSE file.
 
 # {{ ansible_managed }}
-# 2026022701
+# 2026101302
 
 
 # no `set -e` because we change some things at the beginning of the script
@@ -20,6 +20,12 @@ function exit_with_first_failed_retc() {
 }
 
 
+# true if version $1 is the same as or newer than version $2 (plain `X.Y.Z[-N]`, no leading "v")
+function version_ge() {
+    [ "$(printf '%s\n%s\n' "$1" "$2" | sort --version-sort | tail -n 1)" == "$1" ]
+}
+
+
 # make sure the backup is not readable by others
 old_umask=$(umask)
 umask 027
@@ -46,8 +52,10 @@ mysql_table_lock_retc=$?
 
 # https://github.com/mydumper/mydumper
 
-# get installed mydumper version ("mydumper v0.15.1-3, built against MySQL 8.0.33-25 with SSL support")
-mydumper_version=$(mydumper --version | awk '{print $2}' | tr --delete ',')
+# get installed mydumper version ("mydumper v0.15.1-3, built against MySQL 8.0.33-25 with SSL support"
+# on RHEL, "mydumper 0.10.0, built against MySQL 8.0.36" in the Debian and Ubuntu repositories).
+# strip the comma and the leading "v", so that version_ge() compares the same way for both.
+mydumper_version=$(mydumper --version | awk '{print $2}' | tr --delete ',' | sed 's/^v//')
 
 # build the cmdline parameters for mydumper
 mydumper_params="--host=$HOST "
@@ -57,9 +65,8 @@ mydumper_params+="--build-empty-files "
 if [ -n "$COMPRESS" ]; then
     mydumper_params+="--compress=$COMPRESS "
 fi
-if [[ "$(echo -e "$mydumper_version\nv0.15.2-7" | sort --version-sort | tail -n1)" == "$mydumper_version" ]]; then
-    # if $(mydumper --version) > 0.15.2-7:
-    # parameter available since mydumper v0.15.2-8+
+if version_ge "$mydumper_version" 0.15.2-8; then
+    # parameter available since mydumper v0.15.2-8
     mydumper_params+="--dirty "
 fi
 mydumper_params+="--events "
@@ -67,9 +74,18 @@ mydumper_params+="--logfile=$BACKUP_DIR/mydumper.log "
 mydumper_params+="--long-query-guard=$LONG_QUERY_GUARD "
 mydumper_params+="--outputdir=$BACKUP_DIR "
 mydumper_params+="--routines "
+if [ "$THREADS" -eq 0 ] && ! version_ge "$mydumper_version" 0.15.3-1; then
+    # 0 means "number of CPUs" only since mydumper v0.15.3-1; older versions (0.10.x in the
+    # Debian and Ubuntu repositories) wait forever without a single worker
+    THREADS=$(nproc)
+fi
 mydumper_params+="--threads=$THREADS "
 mydumper_params+="--triggers "
-mydumper_params+="--trx-tables=0 "
+if version_ge "$mydumper_version" 0.20.1-1; then
+    # `--trx-tables=0` takes an argument since mydumper v0.20.1-1, where it also turned on by
+    # default. older versions do not know the argument and behave like "0" anyway.
+    mydumper_params+="--trx-tables=0 "
+fi
 mydumper_params+="--verbose=3 "
 mydumper_params+="$RAW "
 
diff --git a/roles/mongodb/README.md b/roles/mongodb/README.md
index d9a96dc5d..7c5417111 100644
--- a/roles/mongodb/README.md
+++ b/roles/mongodb/README.md
@@ -16,6 +16,11 @@ This role is only compatible with the following MongoDB versions:
 *Available since LFOps `2.0.0`.*
 
 
+## How the Role Behaves
+
+* `mongodb-dump.service` runs with `Nice=19`, best-effort IO priority 7 and `CPUWeight=` / `IOWeight=` of `20` (systemd default: `100`), so the dump does not slow down the application on a busy host. The weights only take effect under contention. `IOWeight=` only has an effect on block devices that use the BFQ IO scheduler; with `none` or `mq-deadline`, as on a RHEL 9 VM with virtio disks, it is ignored. The IO class is deliberately not `idle`: the file-based method keeps the instance write-locked while it copies `/var/lib/mongo`, and an idle copy could stretch that lock.
+
+
 ## Known Limitations
 
 * MongoDB publishes no server packages for Debian 13 and Ubuntu 26.04, so the role cannot install MongoDB there.
diff --git a/roles/mongodb/templates/etc/systemd/system/mongodb-dump.service.j2 b/roles/mongodb/templates/etc/systemd/system/mongodb-dump.service.j2
index 5341f0e03..581618390 100644
--- a/roles/mongodb/templates/etc/systemd/system/mongodb-dump.service.j2
+++ b/roles/mongodb/templates/etc/systemd/system/mongodb-dump.service.j2
@@ -1,5 +1,5 @@
 # {{ ansible_managed }}
-# 2022083101
+# 2026101301
 
 [Unit]
 Description=MongoDB Dump Service
@@ -8,6 +8,16 @@ Description=MongoDB Dump Service
 ExecStart=/usr/local/sbin/mongodb-dump
 Type=oneshot
 User=root
+# the dump must not slow down mongod and the applications behind it. Nice= and the CPU / IO
+# weights (default 100) only matter under contention, so an idle host dumps at full speed.
+# best-effort instead of idle: with `mongodb__dump_method_file_based` the instance stays
+# write-locked (fsyncLock) while /var/lib/mongo is copied, and an idle IO class could starve
+# the copy and stretch that lock.
+Nice=19
+IOSchedulingClass=best-effort
+IOSchedulingPriority=7
+CPUWeight=20
+IOWeight=20
 
 [Install]
 WantedBy=basic.target
diff --git a/roles/postgresql_server/README.md b/roles/postgresql_server/README.md
index 9cec62940..a085af59a 100644
--- a/roles/postgresql_server/README.md
+++ b/roles/postgresql_server/README.md
@@ -6,6 +6,11 @@ This role installs and configures a [PostgreSQL](https://www.postgresql.org/) se
 *Available since LFOps `2.0.0`.*
 
 
+## How the Role Behaves
+
+* `postgresql-dump.service` runs with `Nice=19`, best-effort IO priority 7 and `CPUWeight=` / `IOWeight=` of `20` (systemd default: `100`), so the dump does not slow down the application on a busy host. The weights only take effect under contention. `IOWeight=` only has an effect on block devices that use the BFQ IO scheduler; with `none` or `mq-deadline`, as on a RHEL 9 VM with virtio disks, it is ignored.
+
+
 ## Dependent Roles
 
 Any [LFOps playbook](https://github.com/Linuxfabrik/lfops/blob/main/playbooks/README.md) that installs this role runs these for you. Optional ones can be disabled via the playbook's skip variables.
diff --git a/roles/postgresql_server/templates/etc/systemd/system/postgresql-dump.service.j2 b/roles/postgresql_server/templates/etc/systemd/system/postgresql-dump.service.j2
index 0e3c793ff..d71743e86 100644
--- a/roles/postgresql_server/templates/etc/systemd/system/postgresql-dump.service.j2
+++ b/roles/postgresql_server/templates/etc/systemd/system/postgresql-dump.service.j2
@@ -1,5 +1,5 @@
 # {{ ansible_managed }}
-# 2022111801
+# 2026101301
 
 [Unit]
 Description=PostgreSQL Dump Service
@@ -8,6 +8,13 @@ Description=PostgreSQL Dump Service
 ExecStart=/usr/local/sbin/postgresql-dump
 Type=oneshot
 User=root
+# the dump must not slow down PostgreSQL and the applications behind it. Nice= and the CPU / IO
+# weights (default 100) only matter under contention, so an idle host dumps at full speed.
+Nice=19
+IOSchedulingClass=best-effort
+IOSchedulingPriority=7
+CPUWeight=20
+IOWeight=20
 
 [Install]
 WantedBy=basic.target