Hosting Pages
These pages are the same shape: a top-level cross-account view (search + filter) plus a per-account drill-down. On Domains, DNS, SSL, Databases, Email, FTP, Backups and Cron the chosen account filter is part of the address (?account=<username>), so a reload keeps it and a link can open the page already filtered; on the Logs page the Admin Logs account filter is ?tab=admin&username=<username>. The File Manager and the Logs page's Account Logs view have no cross-account list: pick an account first — the File Manager then opens that account's home directory, and Account Logs shows its access and error logs per domain and, when there is one, its PHP error log (the System Logs and Admin Logs views need no account). Operations available match the admin user's full capabilities:
_, $ and -)422 VALIDATION_ERROR. An extract reads the archive that was checked and writes each entry, and sets its owner, through the destination folder's handle: an entry whose name holds a link, a FIFO or another kind of entry at that moment is refused naming the entry (details.field). The folders it makes (a new folder, the missing folders of a new file or upload, an extract's folders) take their mode from the account's home: 0755 where the home keeps other accounts out (every web server's homes, 0750 with the web group), 0750 where other accounts may still pass through it (a LiteSpeed home an earlier release left at 0711 and not yet closed); existing folders and file modes are left as they are. A site's PHP copies a folder's mode onto the folders it makes below it but gives them the account's own group, so under a 0750 folder its files (WordPress media uploads and imported images) answered 403 to every visitor; a site installed that way before this release — on OpenLiteSpeed / LiteSpeed Enterprise too, where the homes were open — is repaired by setting wp-content/uploads to 755 with its sub-items (Permissions)/bin/sh in a login environment and stops it after 85 seconds (the agent’s own limit is 90); a suspended account's crontab is parked until it is resumed; jobs that reached the account's crontab outside the panel (over SSH, by a migration or a restore) are listed and managed too — such a line stays as it is until its job is edited, while settings (MAILTO=), @ shortcuts, named months or days and comments are kept but not listed/var/log/whost-accounts/<user>/ (root's, outside the home) through a handleMulti-select exists only in the File Manager: the selected items are deleted, copied, moved or archived together, and a delete asks a plain confirmation that names the number of items — nothing has to be typed. The other pages act on one row at a time (the check boxes in the Databases, Email and FTP dialogs set privileges and unlimited quota); whole-list actions such as Secure All Domains on the SSL page, the mail queue's Flush Queue and Delete All, and Clean up orphans on the FTP page are single buttons behind an ordinary confirmation.
Backups — scope and settings. A backup holds the account's home files (setuid and setgid bits are not kept: a file or folder that carries one is backed up without it; the PHP session files in the account's tmp folder are left out; the home is read through folder handles without following links, so a folder or file the account turns into a link while the backup runs is left out — counted in the agent log — and devices, FIFOs and sockets are not backed up), one SQL dump per database and the mailboxes (addresses, password hashes, quotas and the mail itself). A restore puts back public_html, tmp, ssl, mail and the folders of the account's addon domains and subdomains that live outside public_html (the sites the account has at the time of the restore; log files, the account's backups folder and hidden files are not restored), drops and re-creates each database from its dump while its users keep exactly the grants they held (a database no user held gets all privileges for the account's users), and re-creates the mailboxes with their mail; a backup whose archive has gone from disk is reported as such (410) and can only be deleted. The restored folders are written through directory handles as the account's user: a link the account placed at one of those names is removed and replaced by a plain folder, and a name that turns into a link while the restore runs makes the restore fail (500, "Failed to restore …") rather than being written through. In a restored mailbox, a link whose target leaves that mailbox (another mailbox or any other place on the server — every mailbox is read by the one mail user) is left out and named in the agent log ("removed symlink escaping mailbox"); the mailbox's own mail comes back as it was, and an archive the agent made itself holds no such link. Each dump is imported by the account's own import user (whost_imp_<account>, a login-locked database user that holds privileges on the account's restored databases only; it is not listed among the account's users and is dropped with the account): a statement in the dump that names another database fails and the restore reports it (500, "Database restore failed for … ERROR 1142 … denied to user 'whost_imp_403 BACKUP_DISABLED and the panel locks the actions), sets the retention window that prunes every backup, and caps the schedules per account (403 BACKUP_SCHEDULE_LIMIT when the cap is reached — the panel shows the reason). Nightly backups for new accounts on the same tab gives every account created from then on a daily schedule (files, databases and mailboxes, kept on this server, with the start time and the number of backups kept you set; an account that already has a schedule gets none); Also add it to existing accounts without a schedule adds it once, when you save, to every account that has none, and the toast names how many got it. It is on for a new installation and off on an installation updated to this version; the setup guide's backup step sets the same switch. The Settings tab also carries Encrypt backup archives (off by default): with it on, every new archive — on-demand or scheduled, for every account — is sealed on the server right after it is made (AES-256-GCM in 1 MiB pieces under a key derived from the server's own encryption secret, the one that also protects remote-destination credentials; the plain tar does not stay on disk), its name ends in .tar.gz.enc, its row carries an Encrypted badge and encrypted: true in the API, and the recorded size and checksum are those of the sealed file. Remote copies and downloads carry the sealed file as it is — it opens only through a restore on the server that made it (another WHost server, or a server whose secret was replaced, refuses it with 409 BACKUP_DECRYPT_FAILED; a server without the secret refuses the restore, and the creation, with 409 BACKUP_ENCRYPTION_KEY_MISSING). Archives made while the switch was off stay as they are and still restore with it on; turning it off leaves the sealed archives sealed. A backup or a schedule can name several remote destinations (up to 10; the create and schedule windows list the server's destinations as rows, any number of them ticked — none ticked reads Local only): the archive is made once and copied to each destination in turn, and the row shows one badge per copy — copied, being copied, or failed with the reason on hover. A remote copy that fails leaves the local archive and the other copies intact; its badge names the failure — also for a copy the agent could not finish because it stopped during the upload ("The remote copy did not finish: the agent stopped during the upload."). A scheduled backup tries each copy up to three times when the failure may not repeat (a dropped, refused or timed-out connection, a cut TLS session, a 5xx / 408 / 429 answer from the provider; 15 and then 45 seconds apart), and stops at once on one that will (credentials, permission, quota, a changed server identity); a Remote Backup Copy Failed notification naming the destination goes out for each copy that failed, after its last attempt. A backup started by hand tries each copy once and answers with a warning when a copy failed. A destination that does not exist or is switched off stops a backup started by hand before any work (404 REMOTE_DESTINATION_NOT_FOUND / 409 REMOTE_DESTINATION_DISABLED); a scheduled run copies to the other destinations and records that one as failed. Deleting a backup — by hand or by a retention rule — removes its copy from every destination it reached, in the store it was copied to; a backup deleted while its copies are still being made stops copying and the copy just finished is removed as well. A destination that a schedule names anywhere in its list cannot be removed (409 REMOTE_DESTINATION_IN_USE). A schedule an account made with its own destinations keeps them when you edit it: the window shows each as a row of its own ("Destination remote_destination_ids; the single remote_destination_id older callers send is still read and joins the list, and every answer carries the first destination in it, the list, and remote_copies (destination, store, state, error and verified per copy — true when the destination reported the archive's own size back after the copy, null when the provider could not read a size; a copy that came back a different size is recorded as failed with both sizes in its reason, and a scheduled backup tries it again as it does a dropped connection). Each manifest row records the archive's size and SHA-256 (size and sha256 in the answer), and a restore first checks the archive against that checksum: an archive that no longer matches is refused with 409 BACKUP_FILE_CORRUPT instead of being unpacked. While a WHost update is being installed, creating a backup or restoring one is refused with 409 UPDATE_IN_PROGRESS (the update ends in a restart of the agent that would cut it) and a scheduled backup that falls due runs right after the update; working folders an interrupted backup left behind are removed in the background after the agent starts. A backup runs inside the request that starts it, so a large account can keep the request open for many minutes; when no answer comes back at all (a gateway timeout, a dropped connection) the panel says the backup may still be running and asks you to refresh the list and check the notifications, rather than reporting a failure — starting it again would make a second copy. A restore answers the same way: when nothing comes back the panel says it may still be running rather than reporting a failure. A refusal the agent names (a disabled system, a limiter) keeps its own message. The long steps — the copy of the home, each database dump, each mailbox copy (also in a restore), the archive and the removal of the working folder — run as long as they keep working, whatever the account's size: a step that reads, writes and computes nothing for 10 minutes is stopped and the backup fails with that step named ("Command did no work for 600s: …"), and no step runs longer than 6 hours.
Remote backup credentials. A destination's credentials are encrypted on disk, and every field that is encrypted comes back masked in the API and the panel — passwords, API keys, OAuth client secrets and refresh tokens alike, in list answers and in the answers to a save. The exception is the Google Drive service-account JSON the admin API still accepts (credentials_json; the panel does not offer it): it is written as plain text to /etc/whost/credentials/gdrive-<id>.json (root, 0600), the destination keeps only that path (credentials_file, returned as it is), and deleting the destination removes the file. An account's own destination cannot name such a file: the client routes answer 422 to credentials_file. Leaving a masked field untouched when editing a destination keeps the stored value; typing over it replaces it. The FTP provider tries TLS first and falls back to plain FTP when the server has none — only for a destination that has never signed in over TLS: once one has, a server that refuses TLS is refused before the password is sent (if the server really dropped TLS, remove the destination and add it again); SFTP goes through the same timeout. The agent checks the server before the password leaves. An FTP server's TLS certificate must verify against the system CA store for the destination's host; a self-signed certificate, or one issued for another name (a shared-hosting server whose certificate carries its own host name), is refused with a message saying which — use the certificate's name as the host, or turn on Accept an unverified certificate for that destination. With the switch on, the certificate of the first successful connection is remembered (its SHA-256 is shown in the edit form) and a different one, on the control or on a data connection, is refused. An SFTP destination remembers the server's host key at its first successful connection (the test before saving, a stored test, or the first copy) and refuses another key before the password is sent; the edit form shows it as ssh-keygen -l prints it. When a server's key or certificate was replaced on purpose, tick Forget it on save in the edit form; a new host or port starts over by itself. A destination on a private, loopback or metadata address is refused (422; the refused ranges are listed in the developer webhooks guide). Every FTP / SFTP connection is opened to the addresses that check approved, so the host name is not looked up again between the check and the connection. The Validated badge is set only by the server's own connection test (Validate before saving, then save within 15 minutes; a changed credential clears it). A disabled destination is refused as a copy target: a manual backup naming it is refused (409), a scheduled one keeps its local archive and records why the copy was not made. A destination a schedule still names cannot be deleted (409, the answer lists the schedules). The Google Drive / Yandex Disk / OneDrive sign-in state is issued by the server and accepted once, on the surface that issued it.
Database access boundaries. Account database users are granted on their own schema only and never hold a global privilege; the API refuses FILE. A user created directly is bound to localhost or one IPv4 address; a wildcard host (%, or a pattern such as 10.0.0.%) is added only through remote database access (below). MariaDB's secure_file_priv is written into the WHost tuning drop-in and confines server-side file reads and writes to one directory (it applies on the next MariaDB restart). Remote database access copies an account's local users to the given host, but MariaDB ships bound to loopback: until you open it to the network deliberately, the panel returns the remote host with a warning that it cannot connect.
Email. A mailbox address is stored in lower case whatever spelling the request carries. A mailbox, forwarder or queue message that does not exist answers 404 (EMAIL_NOT_FOUND, EMAIL_FORWARDER_NOT_FOUND, MAIL_QUEUE_MESSAGE_NOT_FOUND), the plan's mailbox cap 403 EMAIL_LIMIT, a second forwarder to the same destination 409 EMAIL_FORWARDER_EXISTS; two creates that arrive together for one account are handled one after the other. "Unlimited storage" stores quota 0 (the administrator's dialog only: a client may set a mailbox quota between 1 MB and the plan's disk, and the quotas of the account's mailboxes together may not exceed that disk — 422 VALIDATION_ERROR for the range, 403 EMAIL_LIMIT with the allocated / available figures for the sum; a quota the client leaves as it is, is not measured). A save that changes nothing writes no audit row and reloads no service; deactivating a mailbox or changing its password closes its open IMAP/POP3 sessions, and the webmail button refuses a deactivated mailbox (409 EMAIL_INACTIVE). Deleting a mailbox removes its forwarders; the last mailbox of a domain takes the domain row and the domain's remaining aliases with it; removing an addon or parked domain removes the mailboxes and forwarders on it; terminating an account removes them all. Each of these also removes the webmail's own record of the address (the Roundcube user with its identities, address book, collected addresses and settings), so a mailbox created later under the same address — by another account too — starts empty. Suspending an account switches its mailboxes and forwarders off (IMAP/POP3 login, SMTP submission and inbound delivery stop) and unsuspending puts back exactly the rows that were on. The webmail button mints a single-use token; the Roundcube session runs as the Dovecot master user for reading and sending, the system-scope mint (POST /system/webmail/sso) is audited and refuses a mailbox that does not exist or is inactive, and where nginx serves /webmail (nginx and nginx + Apache hosts) the location does not serve Roundcube's schema, scripts, installer or configuration files; on an Apache-only host it refuses only Roundcube's config, temp and logs folders (the installer folder is removed from disk at install). Webmail has no Filters screen: the mail stack carries no ManageSieve server, so the installer leaves Roundcube's managesieve plugin out and the agent switches it off at start on a host without such a server (a host that runs one keeps the plugin). The queue's Delete All sends the required confirm=ALL; a limiter answer on any mail action is shown as the refusal it is; a list whose read failed shows the failure with a Retry button. The Sent Summary lists senders of hosted domains only, counts a message once whatever its recipient count, opens the bounce sender's records from the "(system)" row, and shows log stamps in their real time zone. A mailbox password may be rotated over an HMAC API key (integrations manage mailboxes); the hosting account's own password may not. The mailbox quota is enforced by Dovecot: IMAP reports it and delivery over the limit is refused (quota 0 = unlimited). A forwarder belongs to its mailbox: while the mailbox is inactive the forwarder is off, and the forwarder dialog offers the account's existing mailboxes as the source (a source without a mailbox is refused, 404). A mailbox or forwarder change reloads no service — the maps are live database lookups. Mail submitted on the loopback without SMTP authentication that claims an address of a hosted domain is refused by the hourly-limit policy (554 5.7.1): web applications authenticate as a mailbox or use the sendmail path (PHP mail()), which is not affected. A webmail submission authenticates as the mailbox's master login and spends that mailbox's account allowance. Per-mailbox spam lists (client panel Email → Spam Filter; GET/PUT/DELETE /accounts/{username}/spam/{email} for the administrator) are applied by an Rspamd rule (lua.local.d/whost_per_user.lua) over three recipient-keyed maps the agent rebuilds from the settings files on every save: entries are sender addresses or domains (a domain covers its subdomains), lower-cased, at most 100 per list (422 VALIDATION_ERROR otherwise); a recipient's blacklist rejects the message at delivery (over a server-wide whitelist), its whitelist or filtering switched off accepts the message without scoring unless the server-wide blacklist names the sender; a message to several mailboxes follows the lists only when every recipient's lists agree; the save reports when Rspamd did not reload; a host that stored lists before this rule existed gets the rule and the maps at the agent's next start; a deleted mailbox takes its settings with it. The server-wide whitelist and blacklist (Spam Filter page) keep their entries in lower case whatever spelling is typed, and removing an entry matches it case-insensitively.
Outbound mail limits and DKIM. A plan's hourly email limit is applied by a Postfix policy daemon that the agent installs and keeps running; the check leads the recipient restriction list on port 25 and on the submission/smtps ports alike, so authenticated and webmail submissions are both counted, and it fails open (default_action=DUNNO) if the daemon is ever down. An authenticated submission is counted on the authenticated mailbox's account whatever its envelope sender says, a message is counted once whatever its recipient count, and mail arriving from outside the host is not counted. The allowance is per account, shared across every mailbox, domain and subdomain it owns, and exceeding it defers rather than rejects; the first message deferred in an hour raises the administrator's Email Limit Exceeded notification at the agent's next check (every 15 minutes), at most once per account a day. The daemon is listed on the Services page as WHost Mail Policy. DKIM keys generated for a domain are registered with OpenDKIM at generation time, and existing keys for domains the host still serves are registered on agent start, so a published DKIM record actually matches the signature on outgoing mail. Switching an account's mail DNS on generates the domain's DKIM key and publishes its record with the rest of the mail set; switching it off removes both. OpenDKIM is the only signer: Rspamd's own DKIM/ARC signing is off.
Domain name availability. A hostname is taken as soon as any account serves it — as its primary, addon, parked or subdomain name — and account creation, addon and parked domains all refuse it; a parked alias cannot take a name the same account already serves either. The server's own hostname (and localhost) is reserved: no account may take the name the panel itself answers on. A domain's document root is honoured per domain on every web server, so an addon or subdomain serves its own directory rather than the account's primary docroot; the directory is created with the addon (with a placeholder page) and is kept when the addon is removed. The directory has to resolve inside the account home: a folder name that is a link to a place outside it is refused with 422 VALIDATION_ERROR (a link that stays inside the home, such as www → public_html, still works), the home itself and the panel's private folders (logs, tmp, ssl, backups, mail) are refused with the same error, and the levels the agent creates on the way belong to the account.
Addon domains, subdomains and redirects. Adding an addon domain creates its directory, vhost (on the account's PHP version) and (unless switched off) DNS zone; a parked domain runs the PHP version of the domain it points at; removing it also removes the vhosts of its subdomains, the vhosts and zones of the parked domains pointing at it, and its certificate files. A subdomain that runs a PHP version the account does not use yet gets that version's pool. A redirect rule cannot be moved onto a source another rule of the same domain already uses, and a save that changes nothing is neither rendered nor recorded. Deleting a subdomain removes its vhost and DNS record and renames its own folder directly under the account home to <folder>-removed-<timestamp> (nothing is deleted; a folder inside another site's tree, or one another domain also serves, is left as it is). Renaming a subdomain moves its default folder to the new name through a handle on the account home: a free name, an empty folder or one holding only the placeholder index.html takes it, a folder with content is refused (409), and a link at the new name is refused with 422 VALIDATION_ERROR ("'details.field: subdomain) without being followed. A directory still there under the original name when the name is added again is reused as it is (its files stay the account's). In the addon dialog Auto SSL is off by default — switch it on only for a name that already resolves to this server, since a failed Let's Encrypt order still counts against the quota — and an internationalized name has to be entered in Punycode (xn--…); the dialog says so when accented or non-Latin letters are typed. Editing a subdomain sends only the fields that changed, and "Inherit from account" sets the account's own PHP version. On the Domains, DNS and SSL pages a list whose read was refused (a limiter answer among them) shows the failed read with a Retry button instead of an empty list. Secure All Domains reads the server's domain list page by page, asks for confirmation with the number of domains that have no certificate, and sends them 20 at a time with the progress (sent / total) on the button; the agent takes five such requests a minute, so past a hundred domains the panel waits as long as the agent says and sends the same batch again — several hundred domains take several minutes, and the browser tab has to stay open meanwhile (closing or reloading it stops the remaining batches).
Custom vhost config. The panel has no control for this: custom directives are read, saved and removed through GET / PUT / DELETE /accounts/{username}/domains/{domain}/vhost-config, and only for the account's primary domain — any other domain answers 404 DOMAIN_NOT_FOUND. Saving custom directives re-renders the domain from its stored state, so its redirects, per-domain PHP version and SSL listener survive the save. A config the webserver rejects is undone before the error is returned: the previous vhost file is restored, so one account's bad snippet cannot leave the host in a state where every other account's domain operation fails its config test. A field meant for a web server that is not the provider on the host (for example apache_config on an nginx host) is refused with 422 VALIDATION_ERROR rather than dropped; saves and removals are written to the audit log (vhost_config_updated / vhost_config_deleted).
Databases. A database user is created with the privileges the dialog lists; a request without any privilege is refused (422) instead of being widened to all of them. The collation has to belong to the chosen character set. In the Manage dialog the X next to a user revokes that user's access to this database only (the user keeps its other databases and is dropped once none remain); the confirmation box says so. A remote host is matched literally, so 10.0.0.% can be added while 10.0.0.5 exists and removing a pattern that was never added is refused; removal asks for a confirmation. The create dialog refuses a name or user name that does not start with a letter and a password shorter than eight characters before the request is sent; when the database was created but its user step was refused, the result panel says so and the database is kept. Dropping an owned or orphan schema from the server-wide list also drops the users left without a grant; an orphan schema is deleted whatever the case of its name. The phpMyAdmin button mints a single-use token that carries the MariaDB root login, not a schema — the schema only names where phpMyAdmin opens (none is minted for a schema that does not exist) — and opens the signon in a new tab. phpMyAdmin's Import takes a file of up to 256 MB and may run for up to ten minutes; a larger dump is imported over SSH. The system-scope token (POST /system/phpmyadmin/sso) is minted only for the panel's own MariaDB host (localhost, 127.0.0.1, ::1 or the configured mysql.host); any other host answers 422. The signon endpoints under /sso/* carry no rate limit of their own: they answer only on the loopback with the SSO pin, and a token is single-use with a five-minute lifetime. The roundcubemail and phpmyadmin schemas are not hidden from the admin phpMyAdmin session, which runs as root and sees every tenant schema; a tenant's session is bound to its own database user. To hand phpMyAdmin to an account's customer — a billing system's phpMyAdmin link — GET /accounts/{username}/databases/phpmyadmin-url (API key or admin session; ?db=<short name> opens one database) mints the client panel's kind of session: a single-use token for a MariaDB user that exists for five minutes with grants on that account's databases only, never root. An account without databases answers 404 DATABASE_NOT_FOUND, a db the account does not hold answers the same before any user is created; the Audit log records the token with the scope admin-tenant-ephemeral. A list whose read failed shows the failure with a Retry button rather than "No databases found".
FTP. The directory of an FTP account is stored in one spelling (/public_html/site, never /public_html/./site/) and every level the agent creates below the home directory belongs to the account, so the chrooted session can enter it. A row whose owner column is empty belongs to the account named by its prefix; only rows whose account no longer exists count as orphans for the single and bulk cleanup. The Edit dialog sends only the fields that changed and closes an untouched form without a request, and the server compares as well: a request that carries no field or only the stored values leaves the row as it is and writes no audit row, a new password alone is recorded as a password change (ftp_password_changed), and an update row names only the fields that changed. A limiter answer (429) on create, edit, password change or delete is reported as the refusal it is — no credential card, no success toast. The daemon compares SHA-512 crypt hashes (MYSQLCrypt crypt), requires TLS (TLS 2; a cleartext login is answered 421) and refuses anonymous logins; a host installed before the hash default has to carry MYSQLCrypt crypt in its Pure-FTPd configuration or every panel-created FTP account is refused at login. On AlmaLinux, Rocky Linux and CentOS Stream they live in /etc/pure-ftpd/pure-ftpd.conf (TLS 2, CertFile /etc/pki/pure-ftpd/pure-ftpd.pem, NoAnonymous yes); a RHEL-family host set up by an earlier installer has TLS off and anonymous logins on until those three lines are set and pure-ftpd is restarted. The create dialog suggests the selected domain's document root as the directory (public_html for the primary domain); the FTP user is confined to that directory, and a directory typed by hand is kept. The account's own FTP user, created with the account, is recorded as an ftp_account_created audit row with scope: default and reaches the webhook bridge like any other. Two create requests that arrive together for one account are handled one after the other, so the plan's FTP cap holds. Creating an FTP account requires an active license, as creating a database, an email account or a DNS zone does.
Account logs live outside the homes. Each account's web server, PHP and Node.js / Python app logs are kept in /var/log/whost-accounts/<user>/ (the base root:root 0711, the account's folder root:<user> 0750, the files with the same names as before: <domain>-access.log, <domain>-error.log, php-error.log, <app>-*.log). The web servers, systemd, logrotate and fail2ban open the logs there, never through the account's home. OpenLiteSpeed and LiteSpeed Enterprise workers open the site logs themselves, so while one of them serves the sites the web group may pass through each account's folder — traverse only, an ACL entry for that group (listing and writing stay closed), given at account creation, on the switch and at every agent start, and taken out again on the switch back to Nginx / Apache; processes in the web group can then read the site logs, never write them. The account sees them through logs in its home, a link to that folder: the File Manager opens it read-only, ~/logs over SSH shows the same files, the Logs page and the app log tabs read the folder, and a backup carries the folder as logs/. FTP sessions are confined to the home and cannot open the link: php-error.log and the app logs, which FTP could download while they lived in the home, are read in the panel, the File Manager or over SSH (the web server logs were never readable over FTP). Removing or replacing the link changes only the account's own view; the agent puts it back while the name is free. An existing server moves once, at the first start of an agent with this layout: each log is renamed into the account's folder (open log files keep being written, no line is lost), every generated file that named <home>/logs (vhosts, PHP-FPM pools, app units, gunicorn.conf.py, logrotate rules, the web jail) is rewritten with its original kept under /var/lib/whost/account-logs-move-<stamp>-separate/, and the running web servers and changed PHP-FPM versions are reloaded. A web-scan jail that still watches the logs under their old names in the home is restarted on its own, so it watches every log under its new name; its bans stay (fail2ban restores them, and the HTTP-layer list holds them while the jail restarts). At every start, and when an account or app is created or restored, a domain access log the account could still write (one that came from its home as the account's own file) is given to root with group and other write removed: the web-scan jails ban server-wide on its lines. The account keeps reading it, and the agent log names each such file. An app's own logs (<app>-*.log: a Node.js app's output, gunicorn's access / error logs of a Python app) stay the account's: its process writes them and no jail reads them. The web server writes an app domain's visits to that domain's own logs, which the account cannot write and the web jails and the traffic count read. Going back to an older agent needs the reverse first, with the agent stopped: cd /opt/whost/agent && /opt/whost/venv/bin/python3 -m whost_agent.services.account_logs_move --to home.
PHP. The account's php.ini values (upload and post size, memory limit, execution and input time, max_input_vars, display_errors, error_reporting) are kept in the account record and rendered into every pool of the account from that one place: a version change, a worker-count change or a plan change carries them along, and a new pool starts from the server defaults (Settings › PHP) with the plan's memory ceiling standing in for an unset memory limit. A per-domain configuration writes only the values that differ from the account's into the domain's .user.ini, so an account-level change keeps reaching the domain; a save equal to the account's values writes nothing and Reset (behind a confirmation) removes the file. The agent reads, replaces and removes that file through a handle on the document root, so a link placed as the file or as the document root is never followed: such a file reads as no values, a save answers 422 VALIDATION_ERROR and leaves it as it is, and Reset removes the link itself. A setting the account's PHP-FPM pool fixes with php_admin_value or php_admin_flag — only a hand edit of the pool file does this; the agent writes these values with php_value — is beyond any .user.ini: the domain dialog shows it read-only with the pool's value and a lock note ("Fixed in this account's PHP pool"), the domain answers list it in locked_keys, and a save that asks for another value leaves that key out of the file, still answers 200 and carries a warnings entry naming the key and the value in force (the pool's own value is no change); the audit entry names only the keys written. The spare worker counts of a dynamic pool kept in the account record (php_overrides: pm_start_servers, pm_min_spare, pm_max_spare, whole numbers 1–500) are carried into every render of the account's pools when all three are set and pm_min_spare ≤ pm_start_servers ≤ pm_max_spare ≤ the worker count; otherwise the pool takes the counts derived from the worker count and the agent log names the values it ignored. Changing a domain's version from the row select asks first; the account's version change (on the account page as well as here) keeps the pools that domain overrides still use, rewrites every vhost — parked names included, which follow the domain they point at — and removes pools no domain uses — a subdomain or addon deleted with its own version takes its pool with it. A domain version equal to the account's is stored as inherited, so when the account itself moves onto a domain's override version that override folds into the inheritance and a later account change moves the domain along. upload_max_filesize and post_max_size do not accept "unlimited" (-1); memory_limit does. The ionCube loader is per PHP version and shared by every account on it: the status shown is what the version runs, and a disable that another account still needs leaves the loader in place. Extension toggles and the OPcache clear restart PHP-FPM for every account on that version and ask first; the OPcache dialog saves only when a value changed. Versions are installed and removed in Default PHP Settings (the version list shows 5.6, 7.4 and 8.0–8.5; whether a version is installed is read from its PHP-FPM package, not from the configuration). A version whose packages are not on the host shows "Not installed on this server" and an Install button; after a confirmation (an end-of-life version carries a warning) the server installs it in the background, as the installer does: the PHP repository when the version is not offered yet (Ubuntu: the ondrej/php PPA; AlmaLinux / Rocky / CentOS Stream: Remi), the version's packages (an extension the repository does not build for an old version is left out and named), the PHP-FPM limits (the memory ceiling on Ubuntu, the socket folder and shared pool on the RHEL family, the open file limit), the session defaults drop-in (99-whost-session.ini), the ionCube Loader where the loader exists for that version, and its PHP-FPM service — then the version is offered to accounts. It takes about a minute; the version tile and a card on the PHP page show the step and the percentage, and the job goes on if you leave the page. One install or removal runs at a time and none while other package work runs: an OS package upgrade — one started from Updates › OS Packages or the guarded OS update run ("An OS package upgrade is running. Try again when it finishes.", 409 OS_UPGRADE_IN_PROGRESS) — or a web server plugin's activation, deactivation or update ("A web server plugin is being activated, deactivated or updated. Try again when it finishes.", 409 PLUGIN_TASK_IN_PROGRESS). A guarded run that starts while the job works waits for it to finish (an hour at most); one that started a moment before the job ends it at once, failed, with nothing installed or removed. PHP 5.6 is built for Ubuntu only: on the RHEL family it shows "Not available on this system". A version that is not the default and that no domain or account uses has a Remove button: after a confirmation it is taken off the panel, its PHP-FPM service stops, its packages are removed — the job stops before anything changes when another package would go with them — and what is left of its configuration is moved to /var/lib/whost/php-removed/<version>-<time>/ (nothing is deleted). A removal is also refused while the web server configuration names the version's socket, and the refusal lists the files: phpMyAdmin and webmail run on the PHP version that was the default when they were installed, so that version stays. Every install and removal writes php_version_installed / php_version_removed to the audit log when it starts and when it ends. A version request for a domain the account does not hold answers 404 DOMAIN_NOT_FOUND; the overview, defaults and OPcache reads show a failed read with a Retry button. When a vhost cannot be re-rendered after a version or worker-count change (the webserver configuration test fails), the change is kept and the response carries a warnings entry naming the domain; the site keeps running on its previous configuration until the webserver configuration is repaired. The OPcache save_comments value is written as 1 on every save. The OPcache card shows each version's live counters, asked from that version's PHP-FPM; its settings go to the installer's tuning file (conf.d/99-whost-opcache.ini on Ubuntu) or, on the RHEL family, to Remi's php.d/10-opcache.ini. Disabling a version in the version list only stops offering it to accounts: its FPM service keeps running for the pools that still use it. Every installed version runs with the panel's session defaults (99-whost-session.ini in its PHP-FPM scan directory: PHP's upload progress feature off, strict session ids on; the installer, the panel's version install and the agent's start-up host hardening all write the same file), every pool locks the session serialization format (session.serialize_handler = php), and the web server refuses a request for a sess_<id> file name inside a document root — see the configuration reference, php section. PHP writes each account's sessions to the account's tmp folder; every 30 minutes the agent removes the session files there (sess_*) that have not been written for a day — or for the longer session.gc_maxlifetime a php.ini of the host sets — and the agent log names how many it removed per account. On Ubuntu and Debian neither PHP (session.gc_probability = 0) nor the distribution's phpsessionclean timer expires these folders, so an earlier release left every session a site opened on disk.
DNS zones. A zone the panel creates (account creation, addon or parked domain, the "Create DNS Zone" call to action) carries an SOA naming the configured primary nameserver (Settings › General) and hostmaster.<zone>, and NS records for the configured set; changing the nameservers in Settings rewrites the NS records and SOA primary of every existing zone. Record targets (MX, CNAME, NS, PTR, SRV) are accepted with or without the trailing dot and are shown without it. DNSSEC DS records are listed as SHA-256 and SHA-384 digests only; a save that changes nothing writes no audit row. A zone created from the "Create DNS Zone" call to action is signed with DNSSEC the way a zone created with the account is (a signing failure is reported as a warning and leaves the zone usable). The mail records (MX, mail A, SPF, DMARC and the DKIM key) are one set that follows the account's setup mail DNS switch: a zone bootstrapped while the switch is off — at account creation, for an addon or parked domain, or from the call to action — carries only the A records, and the switch later adds or removes the whole set on the primary zone. On an existing zone the call to action also puts back a missing SOA or apex NS set (a zone without its SOA is answered REFUSED for every name); the SOA and the zone's last NS record are protected — the SOA cannot be edited or deleted and the last NS cannot be deleted until another is added (409 DNS_RECORD_PROTECTED).
Certificates. A Let's Encrypt order validates over HTTP from the panel's own challenge folder, /var/www/letsencrypt/.well-known/acme-challenge — every site's configuration (nginx, Apache and OpenLiteSpeed alike) serves /.well-known/acme-challenge/ from that folder, and nothing is written into the account's site folders — is named after the domain (--cert-name) and installs a renewal hook that refreshes the certificate files the vhost points at and reloads the web server, so the timer's renewals reach the site. When an update changes the site templates or which template a site gets, the first start after it renders every hosted name's site configuration once from the current templates, each with its account's own web server, and every Python / Node.js app's nginx configuration (the previous files are copied to /var/lib/whost/vhost-rerender-<timestamp>/ first, the apps' under apps/; a name or app whose render is refused keeps its previous files and is tried again at the next start); with many sites that first start takes longer. Deleting a Let's Encrypt certificate also removes its lineage from certbot. A custom certificate is refused when it has expired or is not valid yet and when the CA bundle does not parse; the served file carries the chain. The certificate files live under /home/<user>/ssl, which belongs to root so an account cannot remove a file the web server's configuration depends on. That folder has to be a plain folder — when it is a link, installing a certificate answers 422 VALIDATION_ERROR — and each file is written as a new file (key 0600, certificate and chain 0640), so an entry left there beforehand is replaced, not written through. The HTTPS vhost keeps the domain's own document root and redirect rules, and removing a certificate restores the HTTP vhost with them. A certificate is installed into the account's own web server setup: for an account on Apache behind nginx (an account set to Apache on a server where nginx serves the public ports) TLS ends at nginx and the site stays on Apache — its .htaccess rules apply over HTTPS — and installing or removing a certificate reloads Apache together with nginx (a renewal reloads nginx, which holds the certificate); an account on nginx alone is served by nginx and PHP-FPM as before. A forced order for a certificate that is not due yet keeps the current one: the answer says the certificate is current, carries a warning, and no renewal is recorded; force_renew renews it at once. An order covers the name and its www. form, and Let's Encrypt has to reach both over port 80. When it does not issue the certificate, the reason it gave is shown — for each name that failed, what the certificate authority reported (for example www.<domain>: … no valid A records found …, or <address>: Invalid response from http://<domain>/.well-known/acme-challenge/…: 403 when another server or a proxy answers for the name), or for a refused order (a rate limit, a name it does not issue for) certbot's own error line: in the error notice of Install SSL and Renew Certificate, under the Let's Encrypt form of the install dialog until the next attempt, on each failed domain's line of the Secure All Domains result, in the API answer (500 SSL_ERROR: the message and details.reason, one finding per line), in the agent log and so in the error records Report a bug offers. Every 15 minutes the agent reads the certificate each hosted name serves and warns the administrator and the account owner (SSL Expiry Warning) when 14, 7, 3 and 1 days are left — once at each step for that certificate; a renewed or replaced certificate starts over, and a certificate that has already expired is not reported.
Endpoints used: /accounts/{u}/{resource}/... family — see the API reference for each.
Our support team is here around the clock for anything you can't find above.