uncloud/doc/uncloud-manual-2020-08-01.org

17 KiB
Raw Permalink Blame History

Bootstrap / Installation / Deployment

Pre-requisites by operating system

General

To run uncloud you need:

  • ldap development libraries
  • libxml2-dev libxslt-dev
  • gcc / libc headers: for compiling things
  • python3-dev
  • wireguard: wg (for checking keys)

Alpine

apk add openldap-dev postgresql-dev libxml2-dev libxslt-dev gcc python3-dev musl-dev wireguard-tools-wg

Debian/Devuan:

apt install postgresql-server-dev-all

Creating a virtual environment / installing python requirements

Virtual env

To separate uncloud requirements, you can use a python virtual env as follows:

python3 -m venv venv
. ./venv/bin/activate

Then install the requirements

pip install -r requirements.txt

Setting up the the database

Install the database service

The database can run on the same host as uncloud, but can also run a different server. Consult the usual postgresql documentation for a secure configuration.

The database needs to be accessible from all worker nodes.

Alpine
apk add postgresql-server
rc-update add postgresql
rc-service  postgresql start`
Debian/Devuan:
apt install postgresql

Create the database

Due to the use of the JSONField, postgresql is required. To get started, create a database and have it owned by the user that runs uncloud (usually "uncloud"):

bridge:~# su - postgres
bridge:~$ psql
postgres=# create role uncloud login;
postgres=# create database uncloud owner nico;

Creating the schema

python manage.py migrate

Configuring remote access

  • Get a letsencrypt certificate
  • Expose SSL ports
  • Create a user
certbot certonly --standalone \
          -d <yourdbhostname> -m your@email.come \
          --agree-tos  --no-eff-email
  • Configuring postgresql.conf:
listen_addresses = '*'		# what IP address(es) to listen on;
ssl = on
ssl_cert_file = '/etc/postgresql/server.crt'
ssl_key_file = '/etc/postgresql/server.key'
  • Cannot load directly due to permission error:

2020-12-26 13:01:55.235 CET [27805] FATAL: could not load server certificate file "/etc/letsencrypt/live/2a0a-e5c0-0013-0000-9f4b-e619-efe5-a4ac.has-a.name/fullchain.pem": Permission denied

  • hook
bridge:/etc/letsencrypt/renewal-hooks/deploy# cat /etc/letsencrypt/renewal-hooks/deploy/postgresql
#!/bin/sh

umask 0177
export DOMAIN=2a0a-e5c0-0013-0000-9f4b-e619-efe5-a4ac.has-a.name
export DATA_DIR=/etc/postgresql

cp /etc/letsencrypt/live/$DOMAIN/fullchain.pem $DATA_DIR/server.crt
cp /etc/letsencrypt/live/$DOMAIN/privkey.pem   $DATA_DIR/server.key
chown postgres:postgres  $DATA_DIR/server.crt $DATA_DIR/server.key
  • Allowing access with md5 encrypted password encrypted via TLS
hostssl all             all             ::/0			md5
postgres=# create role uncloud password '...';
CREATE ROLE
postgres=# alter role uncloud login ;
ALTER ROLE

Testing the connection:

psql postgresql://uncloud@2a0a-e5c0-0013-0000-9f4b-e619-efe5-a4ac.has-a.name/uncloud?sslmode
=require
g    #+END_SRC

** Bootstrap
- Login via a user so that the user object gets created
- Run the following (replace nicocustomer with the username)
#+BEGIN_SRC sh
python manage.py bootstrap-user --username nicocustomer

Initialise the database

While it is not strictly required to add default values to the database, it might significantly reduce the starting time with uncloud.

To add the default database values run:

# Add local objects
python manage.py db-add-defaults

# Import VAT rates
python manage.py import-vat-rates

Worker nodes

Nodes that realise services (VMHosts, VPNHosts, etc.) need to be accessible from the main node and also need access to the database.

Workers usually should have an "uncloud" user account, even though strictly speaking the username can be any.

WireGuardVPN Server

  • Allow write access to /etc/wireguard for uncloud user
  • Allow sudo access to "ip" and "wg"
chown uncloud /etc/wireguard/
[14:30] vpn-2a0ae5c1200:/etc/sudoers.d# cat uncloud
app ALL=(ALL) NOPASSWD:/sbin/ip
app ALL=(ALL) NOPASSWD:/usr/bin/wg

Typical source code based deployment

  • Deploy using bin/deploy.sh on a remote server
  • Remote server should have

    • postgresql running, accessible via TLS from outside
    • rabbitmq-configured [in progress]

Testing / CLI Access

Access via the commandline (CLI) can be done using curl or httpie. In our examples we will use httpie.

Checkout out the API

http localhost:8000/api/

Authenticate via ldap user in password store

http --auth nicocustomer:$(pass ldap/nicocustomer)  localhost:8000/api/

Database

uncloud clients access the data base from a variety of outside hosts

So the postgresql data base needs to be remotely accessible

Instead of exposing the tcp socket, we make postgresql bind to localhost via IPv6

::1, port 5432

Then we remotely connect to the database server with ssh tunneling

ssh -L5432:localhost:5432 uncloud-database-host

Configuring your database for SSH based remote access

host all all ::1/128 trust

URLs

  • api/ - the rest API

uncloud Products

Product features

  • Dependencies on other products
  • Minimum parameters (min cpu, min ram, etc).
  • Can also realise the dcl vm

    • dualstack vm = VM + IPv4 + SSD
  • Need to have a non-misguiding name for the "bare VM"
  • Should support network boot (?)

VPN

How to add a new VPN Host

Install wireguard to the host
Install uncloud to the host
Add `python manage.py vpn hostname fqdn-of-this-host` to the crontab
Use the CLI to configure one or more VPN Networks for this host

Example of adding a VPN host at ungleich

Create a new dual stack alpine VM
Add it to DNS as vpn-XXX.ungleich.ch
Route a /40 network to its IPv6 address
Install wireguard on it
TODO [C] Enable wireguard on boot
TODO [C] Create a new VPNPool on uncloud with
the network address (selecting from our existing pool)
the network size (/…)
the vpn host that provides the network (selecting the created VM)
the wireguard private key of the vpn host (using wg genkey)
http command

``` http -a nicoschottelius:$(pass ungleich.ch/nico.schottelius@ungleich.ch) http://localhost:8000/admin/vpnpool/ network=2a0a:e5c1:200:: \ network_size=40 subnetwork_size=48 vpn_hostname=vpn-2a0ae5c1200.ungleich.ch wireguard_private_key=… ```

Example http commands / REST calls

creating a new vpn pool

http -a nicoschottelius:$(pass ungleich.ch/nico.schottelius@ungleich.ch) http://localhost:8000/admin/vpnpool/ network_size=40 subnetwork_size=48 network=2a0a:e5c1:200:: vpn_hostname=vpn-2a0ae5c1200.ungleich.ch wireguard_private_key=$(wg genkey)

Creating a new vpn network

Creating a VPN pool

http -a uncloudadmin:$(pass uncloudadmin) https://localhost:8000/v1/admin/vpnpool/ \
network=2a0a:e5c1:200:: network_size=40 subnetwork_size=48 \
vpn_hostname=vpn-2a0ae5c1200.ungleich.ch wireguard_private_key=$(wg genkey)

This will create the VPNPool 2a0a:e5c1:200::/40 from which /48 networks will be used for clients.

VPNPools can only be managed by staff.

Managing VPNNetworks

To request a network as a client, use the following call:

http -a user:$(pass user) https://localhost:8000/v1/net/vpn/ \
network_size=48 \
wireguard_public_key=$(wg genkey | tee privatekey | wg pubkey)
```

VPNNetworks can be managed by all authenticated users.

* Developer Handbook
The following section describe decisions / architecture of
uncloud. These chapters are intended to be read by developers.
** This Documentation
This documentation is written in org-mode. To compile it to
html/pdf, just open emacs and press *C-c C-e l p*.
** Models
*** Bill
 Bills are summarising usage in a specific timeframe. Bills usually
 spawn one month.
*** BillRecord
 Bill records are used to model the usage of one order during the
 timeframe.
*** Order
 Orders register the intent of a user to buy something. They might
 refer to a product. (???)
 Order register the one time price and the recurring price. These
 fields should be treated as immutable. If they need to be modified,
 a new order that replaces the current order should be created.
**** Replacing orders
  If an order is updated, a new order is created and points to the
  old order. The old order stops one second before the new order
  starts.

  If a order has been replaced can be seen by its replaced_by count:
  #+BEGIN_SRC sh
>>> Order.objects.get(id=1).replaced_by.count()
1

Product and Product Children

  • A product describes something a user can buy
  • A product inherits from the uncloud_pay.models.Product model to get basic attributes

Identifiers

Problem description

Identifiers can be integers, strings or other objects. They should be unique.

Approach 1: integers

Integers are somewhat easy to remember, but also include predictable growth, which might allow access to guessed hacking (obivously proper permissions should prevent this).

Approach 2: random uuids

UUIDs are 128 bit integers. Python supports uuid.uuid4() for random uuids.

Approach 3: IPv6 addresses

uncloud heavily depends on IPv6 in the first place. uncloud could use a /48 to identify all objects. Objects that have IPv6 addresses on their own, don't need to draw from the system /48.

Possible Subnetworks

Assuming uncloud uses a /48 to represent all resources.

Network Name Description
2001:db8::/48 uncloud network All identifiers drawn from here
2001:db8:1::/64 VM network Every VM has an IPv6 address in this network
2001:db8:2::/64 Bill network Every bill has an IPv6 address
2001:db8:3::/64 Order network Every order has an IPv6 address
2001:db8:5::/64 Product network Every product (?) has an IPv6 address
2001:db8:4::/64 Disk network Every disk is identified
Tests

[15:47:37] black3.place6:~# rbd create -s 10G ssd/2a0a:e5c0:1::8

Decision

We use integers, because they are easy.

Distributing/Dispatching/Orchestrating

Variant 1: using cdist

  • The uncloud server can git commit things
  • The uncloud server loads cdist and configures the server
  • Advantages

    • Fully integrated into normal flow
  • Disadvantage

    • web frontend has access to more data than it needs
    • On compromise of the machine, more data leaks
    • Some cdist usual delay

Variant 2: via celery

  • The uncloud server dispatches via celery
  • Every decentral node also runs celery/connects to the broker
  • Summary brokers:

    • If local only celery -> good to use redis - Broker
    • If remote: probably better to use rabbitmq
  • redis

    • simpler
  • rabbitmq

  • Advantage

    • Very python / django integrated
    • Rather instant
  • Disadvantages

    • Every decentral node needs to have the uncloud code available
    • Decentral nodes might need to access the database

      • Tasks can probably be written to work without that (i.e. only strings/bytes)
log/tests

(venv) [19:54] vpn-2a0ae5c1200:~/uncloud$ celery -A uncloud -b redis://bridge.place7.ungleich.ch worker -n worker1@%h logfile ~/celery.log - Q vpn-2a0ae5c1200.ungleich.ch

Variant 3: dedicated cdist instance via message broker

  • A separate VM/machine
  • Has Checkout of ~/.cdist
  • Has cdist checkout
  • Tiny API for management
  • Not directly web accessible
  • "cdist" queue

Milestones   uncloud

1.1 (cleanup 1)

TODO [C] Unify ValidationError, FieldError - define proper Exception
  • What do we use for model errors
TODO [C] Cleanup the results handling in celery
  • Remove the results broker?
  • Setup app to ignore results?
  • Actually use results?

1.0 (initial release)

TODO [C] Initial Generic product support
  • Product
TODO [C] Recurring product support
TODO [C] Support replacing orders for updates
DONE [A] Finish split of bill creation

CLOSED: [2020-09-11 Fri 23:19]

TODO [C] Test the new functions in the Order class
Define the correct order replacement logic

Assumption:

  • recurringperiods are 30days
Case 1: downgrading
  • User commits to 10 CHF for 30 days
  • Wants to downgrade after 15 days to 5 CHF product
  • Expected result:

    • order 1: 10 CHF until +30days
    • order 2: 5 CHF starting 30days + 1s
    • Sum of the two orders is 15 CHF
  • Question is

    • when is the VM shutdown?

        1. instantly
        1. at the end of the cycle
    • best solution

      • user can choose between a … b any time
Duration
  • You cannot cancel the duration
  • You can upgrade and with that cancel the duration
  • The idea of a duration is that you commit for it
  • If you want to commit lower (daily basis for instance) you have higher per period prices
Case X
  • User has VM with 2 Core / 2 GB RAM
  • User modifies with to 1 core / 3 GB RAM
  • We treat it as down/upgrade independent of the modifications
Case 2: upgrading after 1 day
  • committed for 30 days
  • upgrade after 1 day
  • so first order will be charged for 1/30ths
Case 2: upgrading
  • User commits to 10 CHF for 30 days
  • Wants to upgrade after 15 days to 20 CHF product
  • Order 1 : 1 VM with 2 Core / 2 GB / 10 SSD 10 CHF

    • 30days period, stopped after 15, so quantity is 0.5 = 5 CHF
  • Order 2 : 1 VM with 2 Core / 6 GB / 10 SSD 20 CHF

    • after 15 days
    • VM is upgraded instantly
    • Expected result:

      • order 1: 10 CHF until +15days = 0.5 units = 5 CHF
      • order 2: 20 CHF starting 15days + 1s … +30 days after the 15 days -> 45 days = 1 unit = 20 CHF
      • Total on bill: 25 CHF
Case 2: upgrading
  • User commits to 10 CHF for 30 days
  • Wants to upgrade after 15 days to 20 CHF product
  • Expected result:

    • order 1: 10 CHF until +30days = 1 units = 10 CHF
    • order 2: 20 CHF starting 15days + 1s = 1 unit = 20 CHF
    • Total on bill: 30 CHF
TODO [C] Note: ending date not set if replaced by default (implicit!)
  • Should the new order modify the old order on save()?
DONE Fix totally wrong bill dates in our test case

CLOSED: [2020-09-09 Wed 01:00]

  • 2020 used instead of 2019
  • Was due to existing test data …
DONE Bill logic is still wrong

CLOSED: [2020-11-05 Thu 18:58]

  • Bill starting_date is the date of the first order
  • However first encountered order does not have to be the earliest in the bill!
  • Bills should not have a duration
  • Bills should only have a (unique) issue date
  • We charge based on bill_records

    • Last time charged issue date of the bill OR earliest date after that
    • Every bill generation checks all (relevant) orders

      • add a flag "not_for_billing" or "closed"
      • query on that flag
      • verify it every time
TODO Generating bill for admins/staff
Bill fixes needed
TODO Double bill in bill id
TODO Name the currency
TODO Maybe remove the chromium pdf rendering artefacts
  • date on the top
  • title on the top
  • filename bottom left
  • page number could even stay
TODO Try to shorten the timestamp (remove time zone?)
TODO Bill date might be required
TODO Total and VAT are empty
TODO Line below detail/ heading