Contact

Subscribe via Email

Subscribe via RSS/JSON

Categories

Recent Posts

Creative Commons Attribution 4.0 International License
© Rakhesh Sasidharan

Elsewhere

Notes on Dynamic DNS updates, DNS Scavenging, etc.

Dynamic DNS updates can be set to one of these (per zone):

  • None => No dynamic updates are allowed for the zone on this server
  • Secure => Only secure updates are allowed.
    • Note: This is only applicable to AD integrated zones.
    • By default only domain members (domain joined computers & domain users) are allowed to update the zone for secure updates. This is controlled by the ACLs on the zone (which can be viewed via the Security tab of the zone – check out the ACE for “Authenticated Users“). See this link for more
  • Nonsecure and secure => Both secure and nonsecure updates are allowed.

Scavenging

Dynamic DNS updates result in records being added and deleted to DNS. But while records are correctly added, it is not always the case that the record is also correctly removed. For instance, a client could have got an IP address from DHCP and dynamically registered its A record. Maybe the client then crashed so it never removed the dynamically registered record. The address, however, is removed from DHCP after the lease expires and could later be assigned to another client – who also dynamically registers itself in DNS – resulting in two A records, both to the same IP address, but one of them incorrect. To prevent such issues DNS scavenging is required. This removes stale DNS records after a pre-defined period. 

  • Scavenging is set at 3 places on a Windows server and all three must coincide for a record to be scavenged. These places are:
    • an individual record; 
    • the zone; and
    • the server performing the scavenging. 
  • The scavenging setting on an individual record can be viewed only after selecting View > Advanced in the DNS MMC and then viewing the properties of a record. 
    • When a dynamic DNS record is created it has a timestamp (rounded down to the nearest hour when the record was created).
      • When a record is first created it is considered an “Update”.
      • When an existing record is updated with the same IP address it is considered a “Refresh”.
      • When an existing record is updated with a new IP address it is considered an “Update”.
    • Every 24 hours Windows clients will attempt to to dynamically update the DNS record. The update could be considered an update or a refresh depending on whether the IP changes as above.
    • If a record is enabled for scavenging its properties window will have a tick next to “Delete this record when it becomes stale”. 
      • Static records don’t have this ticked by default (because they are not meant to be scavenged). 
      • If this is manually ticked (for a static or dynamic record) then a timestamp will be set to when it was ticked (rounded down to the nearest hour). 
    • It is possible to set scavenging on a zone and all its records via the following command: dnscmd /ageallrecords
      • This is not recommended as it enables scavenging on all records – even static. Do not use this command on zones with static records. 
  • The scavenging setting for a zone can be viewed via the Aging button in the zone properties (by default the setting is off).
    • The aging values for a zone are replicated to all DNS servers hosting the zone.
    • Two intervals are in play here:
      • No-refresh interval: Once a record is refreshed, it is not refreshed again until this time period has passed. 
        • The purpose of this seems to be to reduce replication traffic. If a client refreshes its DNS record every 24 hours, those are ignored by the DNS server for the no-refresh interval, and not replicated to other DNS servers.  
      • Refresh interval: How much time to wait once a record is refreshed before it can be scavenged?
        • So this interval specifies how long the server should wait after a record has refreshed before it can considered it a candidate for scavenging.
        • The default value is 7 days. This means, if a record is refreshed today, the server will wait for 7 days to see if it’s refreshed again. If it is not, the server considers this record ready for scavenging. 
    • Both intervals must be passed for a record to be expired. By default both are 7 days, so what this means is:
      • If a record is created/ updated/ refreshed today, for the next 7 days the record is considered current – irrespective of whether any refreshes happen or not (because remember: during the no-refresh interval refreshes, if they happen, are ignored so the server considers the record as current for this period). 
      • After those 7 days have passed, the server checks if there are any refreshes.
        • If there are, the timestamp is accordingly updated and it goes back into waiting the no-refresh interval again. 
        • If there are no refreshes, the server now waits 7 days of refresh interval to see if any refreshes happen. If they do, the record goes back into the no-refresh interval; if there aren’t any, the record is ready for scavenging. 
    • As an aside, the default lease duration for Windows server DHCP leases is 8 days. Which is why the no-refresh interval is set to 7 days by default. During these 7 days the address won’t be allocated to any other client, nor will it change with the client, so chances of a refresh are minimal. 
      • DHCP leases and Dynamic DNS updates can conflict if clients are responsible for updating DNS with their addresses (which is usually the default).
      • Say a client got an IP address from DHCP (leased to it for 8 days, remember). The client will update that in DNS. For the next 7 days any refreshes from the client are ignored (no-fresh interval, expected). From the 7th day any refreshes/ updates will be considered.
      • Say our client went offline on the 3rd day. So on day 7 it doesn’t send a refresh – no problemo, DNS will not scavenge the record yet, it will simply wait for another 7 days.
      • On the 8th day, however, DHCP will release that IP address for others to use. Any new client that comes up will now get this address. This new client will send a Dynamic DNS update to the DNS server – creating a new A record to the same address, but with this new client’s name. Thus there are two DNS entries now to the same IP address!
      • Only after the refresh interval expires (7 days) can the old record be actually scavenged by the server (and even then there could be a delay based on the server setting – see below). 
      • For this reason it is recommended that the DHCP lease duration match the “no-refresh+refresh” interval of DNS scavenging. In the default case, either increase the DHCP lease to 14 days (7+7 days) or decrease the no-refresh and refresh intervals to 4 days (so the sum is 8 days, the DHCP lease).
        • Alternatively, allow the DHCP server to make updates on behalf of clients and disable (via GPO?) clients from registering updates with the DNS server.
          • Read this post on DHCP servers and Dynamic DNS updates.
          • Typical solution is to put all DHCP servers in a group called DnsUpdateProxy, but that’s not recommended – especially for DCs – because if a server is in this group the dynamic DNS records it creates have no security (so in the case of a DC this means the SRV records written by netlogon can be changed by anyone!)
          • It is better to create a low privilege AD user and get all DHCP users to use that account to register records. This way the dynamic DNS records are secured to that user.
          • Also note: if dynamic DNS records are written by a server in the DnsUpdateProxy group – i.e. with no security – if any other machine (even one not in this group) changes this record (because the records are open to all) the ACLs of that record will be changed to only grant that machine permissions to the record. Thus the original DHCP server will lose rights to that record. DnsUpdateProxy is not a good idea. 
        • It is important that clients be disabled from registering dynamic DNS updates in this case. Else the ACLs on the DNS record created by the client will prevent updates/ deletions from the DHCP server to the DNS server for those records.  
    • When scavenging is enabled for a zone, the “Date and time” the zone can be scavenged after value is set to the time the setting was enabled (rounded down to the nearest hour) plus the refresh interval period. 
    • It is also possible to right click a DNS server and set scavenging values for all zones on that server. These only apply to zones created after this setting was changed (unless the setting to modify existing zones is explicitly selected). 
  • The scavenging setting on the server can be enabled via the “Advanced” tab of the server properties in the DNS MMC (by default the setting is off).
    • When this setting is enabled, the scavenging period is set to a default of 7 days. The scavenging period defines how often the server will try to scavenge records. 
      • Does this mean every time the server starts scavenging all records are immediately deleted? No – because you have to also consider the no-refresh and refresh intervals of above. When a server runs its scavenging task, if a record to be scavenged has not crossed the refresh interval, it will not be removed. Similarly, if a record has crossed its refresh interval and is ready to be scavenged, if the server’s scavenging period isn’t due for a few more days nothing will happen. It’s only when the scavenging period is due that this record will be scavenged. 
    • When the server scavenges records it logs an event ID 2501 indicating how many records were scavenged. If no records were scavenged, an event ID 2502 will be logged. 
    • Note: You needn’t enable the scavenging setting on all servers hosting a zone. As long as any one server scavenges, the changes will propagate to others. In fact, it’s preferred to have only one server (or a set of servers) scavenge a zone as that will make it easier to troubleshoot. If all servers hosting a zone have scavenging on and the zone records are not being scavenged, we will have to check all these servers to see why scavenging isn’t happening. 
    • In practice, it is likely that all servers have scavenging turned on (because they are hosting multiple zones and could be responsible for scavenging one of those zones). But once a server has scavenging turned on it will scavenge any zones that has scavenging turned on. It is possible to restrict the servers that are allowed to scavenge a zone – even if the server and zone have scavenging turned on – via the dnscmd command. The syntax is as follows:

      The IP addresses are optional. If no address is given, all servers are allowed to scavenge it. Example:

      To see what servers have permissions to scavenge a zone the same command with a different switch can be used:

      Resetting this is simple – just don’t specify an IP addresses, that’s all:

The latter part of this blog post gives an example of how scavenging works with all the intervals above. In fact the whole blog post is worth a read.

Quickly rename a bunch of files via PowerShell

Each time I do PowerShell now, it saddens me. I haven’t used it much for the past year I am slowly forgetting even the basic stuff. I really must do some PowerShell coding project soon just to amp up my skills. 

Anyhoo, I needed to rename a bunch of files quickly today. Used the following to rename them:

The files were of the format “The Title (digital) Extra Stuff.cbz“. I wanted to rename it such that the new file name is only the underlined parts. Above code does that to a directory full of such files. 

SOA records (and dynamic DNS in Windows)

I am on the DNS section of my notes from the AD WorkshopPLUS I attended a few months back. That’s why the recent posts are about DNS …

The SOA (Start of Authority) record is something DNS administrators are familiar with. It specifies details about the zone such as the serial number (which can be used by secondary name servers to know the zone has changed), the preferred refresh periods for secondary name servers to sync the zone, the time between retries, whom to contact, the primary name server, and so on. Here’s the SOA record for my rakhesh.com domain:

(In the example above the results also include all the name server records of the zone, but that needn’t be the case always).

In traditional zones you have one primary name server and many secondaries. So you can set one server as the primary in the record above. But what about AD-integrated zones? Since each DNS server is also a primary in that case, things are a bit different. 

What happens is that the primary name server is set to the name of whichever DNS server you ask. Thus, if you query WIN-DC01 for the SOA record to rakhesh.local, it will give itself as the primary, while if you query WIN-DC02 it will return itself as the primary. 

In Windows the name server returned by the SOA record is also used by clients for dynamic DNS updates. Clients query DNS for the SOA record. Whichever server they get a response from will return an SOA record containing itself as the primary name server. Clients then use that name server to dynamically register their A and PTR records. 

An exception to the above is Read-Only DCs (RODCs). These point to another server as the SOA for the zone. A new server is selected every 20 mins. When clients contact a RODC DNS server, they thus get another server as primary in the SOA record and send their dynamic updates to this other server. 

PortQry and PortQryUI

I was aware of PortQry but didn’t know it has a GUI counterpart too PortQryUI. For a quick read on PortQry check out this link, if you have more time and interest check out this KB article. PortQry/ PortQryUI can be used to check the status of TCP and UDP ports on a remote computer. For TCP ports I usually do a telnet to the port (by habit) but didn’t have any equivalent tool for UDP ports. 

Important PortQry switches (as a reference to myself) are:

  • -n -> specifies the server name/ IP address to target
  • -p -> specifies the protocol to test (options are tcp or udp or both; default is tcp)
  • -e -> specifies the port(s) to test (default is port 80)
    • note: it is possible to specify a single port (e.g. -e 81), ports (e.g. -e 80,81) or a range (e.g. -e 80-1024)
    • note: the ports must be in the range 1-65535
  • instead of -e you can use either of the following too:
    • -r -> specifies a port range (e.g. –r 80:90)
    • -o -> specifies a comma-separated list of ports to check in order (e.g. -o 80,443,139)

Some other switches are:

  • -nr -> stops PortQry from resolving an IP address to a name
  • -sl -> waits longer for replies from UDP systems (sl == slow link)
  • -l -> specifies a log file to output to
    • -y -> will over-write the log file if it exists, without prompting

While writing this post I learnt that PortQry can also enumerate the local ports. Nice!

  • The -local switch will list all active TCP/UDP ports on the local system. (Think of it as netstat -a but without any details of the remote end).
  • The -wport (port number) switch will watch a specified port’s state and report when it changes
    • This didn’t work for me, got an error “Port to process mapping is not supported on this system”.
  • The -wpid (PID) switch will watch a specified process ID (PID) and reports when its state changes
    • This too didn’t work for me, same error as above.

A good thing about PortQry is that it can also query protocols that it’s aware of. Thus, for instance, if you query port 53/ UDP (DNS) and something’s listening at the remote end, PortQry can send an additional DNS query to that port. 

This is useful in AD troubleshooting too. For instance, to check whether port 389 of a DC has an LDAP server listening as it should be:

Similarly, RPC:

In the output above, for instance, I query port 135/ TCP which is where the RPC end-mapper service listens. After querying this port and getting a response, PortQry asks it to enumerate the listening services. Of these UUID 12345887-... is what the netlogon service registers under, which as we can see from the above output is listening on ports 49155 (via TCP), 49158 (via TCP), and 49157 (via HTTP). (Netlogon registers with RPC and uses dynamic ports as we saw above, so querying the RPC end-mapper service is the only way to find what ports Netlogon is listening on). 

In contrast to PortQry, PortQryUI has options to query for the services it is aware of. So, for instance, one can use it to query the “Domains and Trusts” service on a DC and it will do PortQry queries to port 135/TCP, port 389/BOTH, port 445/TCP, port 137/UDP, and a few other AD related ports and emit the output in a window (you can see part of the output in the screenshot below). 

portqryUI

SRV records and AD

Example of an SRV record:

Format is:

Where:

  • class is always IN
  • TTL is in seconds 
  • service is the name of the service (example: LDAP, GC, DC, Kerberos). It is preceded by an underscore (prevents collisions with any existing names). 
  • protocol is the protocol over which the service is offered (example: TCP, UDP). It is preceded by an underscore (prevents collisions with any existing names). 
  • domain is the DNS name of the domain for which the service is offered. 
  • host is the machine that will provide this service. 
  • SRV is the text SRV. Indicates that this is an SRV record. 
  • priority is the priority of the host. Similar to SMTP MX record priorities. Host with lower number has higher preference (similar to SMTP records). 
  • weight allows for load balancing between hosts of the same priority. Host with higher number has higher preference (intuitive: higher weight wins).

Similar to MX records, the host of an SRV record must be a name with an A or AAAA record. It cannot point to a CNAME record. 

A machine starting up in a domain can easily query for DCs of that domain via DNS. DCs provide the LDAP service, so a query for _ldap._tcp.dnsdomain will return a list of DCs the machine can use. 

On DCs, the netlogon service registers many records relevant to the services offered by the DC. The A records of the DC are registered by the DNS client service (for Server 2008 and above) but the other records are taken care of by the netlogon service. A copy of the registered records is also stored in the %systemroot%\system32\config\netlogon.dns file in case it needs to be imported manually or compared for troubleshooting. Note that each DC only adds the records for itself. That is, WIN-DC01 for instance, will add a record like this:

It will not add records for WIN-DC02 and others as it does not know of them. The records added by each DC will be combined together in the DNS database as it replicates and gets records from all the DCs, so when a client queries for the above record, for instance, it will get records added by all DCs. 

AD creates a sub-domain called _msdcs.domainname to hold items pertaining to AD for that domain. (MSDCS = Microsoft Domain Controller Services) This sub-domain is created for each domain (including child-domains) of the forest. The _msdcs sub-domain belonging to the forest root domain is special in that it is a separate zone that is replicated to all DNS servers in the forest. The other _msdcs sub-domains are part of the parent zone itself. For instance below are the _msdcs sub-domains for a forest root domain (rakhesh.local) and the _msdcs sub-domain for a child domain (anoushka.rakhesh.local). Notice how the former is a separate zone with a delegation to it, while the latter is a part of the parent zone. 

For forest-root domain

For forest-root domain

For child-domain

For child-domain

Under the _msdcs sub-domain a convention such as the following is used:

Here DcType is one of dc (domain controller), gc (global catalog), pdc (primary domain controller), or domains (GUIDs of the domains).

The various SRV records registered by netlogon can be found at this link. Note that SRV records are created under both the domain/ child-domain and the forest root domain (the table in the link marks these accordingly). Below are some of the entries added by netlogon for DCs – WIN-DC04 and WIN-DC05 – in a site called KOTTAYAM for my domain anoushka.rakhesh.local. Of these WIN-DC05 is also a GC. 

Advertise that the DCs offer the LDAP service over TCP protocol for the anoushka.rakhesh.local domain:

Advertise that the DCs offer the LDAP service over TCP protocol for the anoushka.rakhesh.local domain for both sites of the forest (even though the DCs themselves are only in a single site – this way clients in any site can get these DC names when querying DNS SRV records):

WIN-DC05 has a few additional records compared to WIN-DC04 because it is a GC. 

Notice all of them are specific to its site and are created in the forest root domain zone/ _msdcs sub-domain of the forest root domain. This is because GCs are used forest-wide also. In contrast, similar records advertising the DC service are created for both sites and in the _msdcs sub-domain of the child zone:

To re-register SRV records, either restart the netlogon service of a DC or use nltest as below:

 

Useful ldp queries

Just dumping useful ldp.exe queries as I come across/ think of them. Better to keep them here in one consolidated post for future reference …

Find all GCs in the forest

If a DC is a GC, its NTDS Setting object has an attribute called options whose value is 0x1. Not sure if a DC is a GC and something else, whether the value changes, but for now I’ll assume it doesn’t and so one can quickly search for all GCs in the forest by connecting to the Configuration partition and filtering by the following:

  • Filter: (&(cn=NTDS Settings)(options=1))
  • Base DN: CN=Configuration,DC=domainname
  • Attributes: distinguishedName
  • Scope: Subtree

 

Quick ffmpeg tricks (extract audio, convert audio codec of a video)

I use ffmpeg to convert between audio and video formats. Two situations I use it for are: (1) when I have a video whose audio codec is of a format my media players can’t handle (I use WD TV Live and it has trouble with certain formats), or (2) when I want to extract the audio only from a video. 

For converting only the audio here’s the command I use:

The syntax is pretty obvious. The input file is taken, the video codec is copied as-is to the new file, the audio codec is re-encoded to mp3. I could have used -acodec mp3 too. In the past there used to be an in-built mp3 encoder as well as the mp3 encoder provided by the LAME project (ffmpeg must be built with LAME encoder via the --enable-libmp3lame switch for this to work) so you could choose between either via the two switches, but now there’s no in-built encoder so both --acodec mp3 and --acodec libmp3lame do the same. 

To confirm what switches ffmpeg was built with simply run the command. For example, on my machine:

Notice it was built with --enable-libmp3lame

When it comes to extracting just the audio from a video there’s two ways to do it: 1) you can simply extract the audio in the codec as it is, or 2) you can extract & convert to the codec you want. The latter has the disadvantage that if the original video is in a lossy audio codec, converting will result in some degradation of quality. 

To check what audio codec the file is in, do the following: 

Notice it identifies the audio stream as AAC in this case. If I am happy with extracting that as it is I can do the following:

The -vn switch tells ffmpeg to ignore the video. The -acodec copy switch tells it to copy the audio codec as it is. Since this is an AAC file, I assign an extension of .aac to the output file. 

However, if I didn’t want the audio stream as AAC, I would have done the following:

Here I am converting the audio to mp3. Once again I ignore the video via -vn. I specify the audio codec via -acodec libmp3lame. The rest of the switches are as follows:

  • -ac 2 => two channels (stereo) (note this is same as the input, so is an optional switch)
  • -ar 44100 => sammple rate 44100 Hz (CD quality) (note this is same as the input, so is an optional switch)
  • -ab 320k => bit-rate of 320 kb/s (if I don’t specify this the bit-rate will be 128 kb/s for mp3)

Essentially, instead of just copying the audio codec you convert it. Otherwise the idea is the same. Apart from libmp3lame (or mp3) I could have also used the following audio codecs: vorbis (for ogg), aac, flac, and wma

Before concluding, here’s a link to ffmpeg’s documentation (for all the command-line switches etc). Also, this is a good page on audio/ video containers and ffmpeg. The latter is a very thorough and informative page, I am sure I’ll be referring to it in the future. 

[Aside] A great discussion on copyright and patents

Check out this episode of the Exponent podcast by Ben Thompson and James Allworth.

Ben Thompson is the author of the stratechery blog which is an amazing read for Ben’s insights into technology. James Allworth, I am not much aware of but he was terrific in the above episode. I began following this podcast recently and have only heard one episode prior to this (on Xaomi, again a great episode; in fact the one on copyrights and patents continues on a point mentioned in the Xaomi episode).

This episode was great for a couple of reasons. I felt Ben was caught out of element in the episode, which is unlike how I have read/ heard him previously where he is confident and authoritative. In this episode he was against abolishing copyrights – which is what James was advocating for – but he didn’t have convincing arguments. So he resorted to the usual arguing tricks like prop up examples and try to get the argument to be about the example (and when it still failed he would withdraw the example claiming it wasn’t appropriate here). Or he’d just take a firm stand and refuse to budge. Or incite James by insults and such. Or try and win by conflating the argument with something else which had no relation to it. Basically, usual debating tricks I believe, but it was fun to hear and I was surprised to hear him resorting to these.

Eventually when Ben clarified his point it made sense to me. His argument is that patents are harmful when they apply to “ingredients” (parts of an invention, e.g. pull to refresh) but he has no issues when it applies to a whole thing (e.g. medicine). Moreover, the question is whether the presence of the patent is required to spur invention (not required in the case of technology, required/ preferred in case of medicines) and whether society would be better off without the monopoly afforded by patents (again, no in the case of tech as it leads to barriers of enter and unnecessary patent wars and trolling for new inventions). Copyright, for Ben, is neither harmful to society nor will its absence spur more innovation, so he doesn’t see why it must be abolished. He seems to agree that copyright has its negatives and is harmful in some cases, but he still feels it is useful to make supply scarce (by preventing others from copying the work).

James agrees with most of these but his point is that the same effect can be arrived at without copyrights. Maybe by innovation in other areas, or by agreements between the creator and audience. His argument is more about considering a world without as an option, and to look at how things can be done differently. Moreover, such a world will lead to more creativity and he feels that’s better in the long run.

I can’t write more as I have a flight to catch, so I’ll end this post now. And it’s better to hear the arguments than my summary. Go check out the podcast. It’s a great one! Skip the first few minutes as it is some user feedback etc.

[Aside] Improving your PuTTY connections

Wish I had come across this during my PuTTY days!

The TL;DR summary is that by default PuTTY uses an ASCII encoding while most Linux and other OSes uses UTF-8 encoding. It’s because of this mismatch that manpages and such sometimes appear with â characters. Change the PuTTY encoding and find happiness! 

Start Menu and other directory junctions access denied

Here’s a screenshot of my c:\Users\UserName folder. Notice the “Start Menu” folder:

users-folder

It is a directory junction to another folder, left behind for compatibility purposes. If I double click the folder though I get an access denied:

access-denied

 

The actual target to which this points is accessible, but the junction itself isn’t. Every time I encounter this I think I must make a note of “why” this is so, but I forget. Today I intend to note it down once and for all. 

Remember from an earlier post: there are hard links and there are soft links. The former points to the data itself, and so only work for files. The latter points to the folder/ file containing the data. Soft links are an evolution of directory junctions (there are directory junctions and volume junctions). While directory junctions make use of something called reparse points and were introduced in Windows 2000, soft links were introduced in Vista and are baked into the kernel itself. Microsoft uses directory junctions – as we saw above – to redirect some of its special folders.

The important thing with both directory junctions and soft links is that they can have their own ACLs. So while a user might have full permissions to the target folder, the directory junction or soft link itself may not grant the user permissions and so the contents cannot be accessed via the directory junction or soft link. That’s what happening here too. 

First, from the command prompt note that this is a junction and that I can enter the directory junction but cannot see any files:

The files exist, however, as I can directly browse the target:

Compare the ACLs of the target and directory junction and we see the problem:

(I could have used the Explorer GUI here but I prefer icacls. In the GUI we have to dig down a bit more to see the relevant ACEs).

Notice a DENY entry for Everyone on the directory junction for listing the contents (RD). That’s why I can’t list the junction contents (in Explorer double clicking results in trying to list the contents, while in Command Prompt entering a junction and listing are two separate tasks – that’s why entering the junction worked, but listing the contents failed). 

What the above tells us is that only listing the junction contents is prohibited. If we know the names of some folders in there – as older software for whom this directory junction is present would know – we can go to those locations using the directory junction. Thus the following works:

There you go!

Add Readability and Instapaper ‘Read Now’ links to your posts

Yesterday I went back and read one of my older posts from my tablet. That made me realize my blog doesn’t have a mobile friendly view. Sure, it doesn’t have any ads or widgets, and the posts appear clean on a browser as my emphasis is on the text/ code, but that doesn’t translate well to a mobile device as the fonts are small and a bit of zooming and scrolling is required to hide the left sidebar and other bits. 

Eventually I read the post using a Readability bookmarklet I had on the mobile browser so that got me thinking I must add quick links to do this for each post so any visitors can take advantage of the same. 

I use Instapaper for most of my reading. For some posts that Instapaper has difficulty rendering (mostly posts with a lot of code, pictures) I use Pocket. I prefer Readability over Pocket as its iOS app is terrific, but Readability’s Android app sucks (poor UI, syncing issues, doesn’t keep track of my last read location) and so I use Pocket rather than Readability. 

For publishers Readability offers an embed code. This is good in that it allows one to read the page in a Readability view without adding to Readability (similar to what I did yesterday using the bookmarklet). It also lets you add the page to Readability, print it, send to Kindle, or email – useful stuff. What I don’t like about the embed code, though, is that it pulls in JavaScript from their website and adds a block of buttons to my posts. I don’t want a block of buttons – in my case, all I want is to offer users a link they can click to get the page in a Readability view. 

For readers Readability offers the bookmarklets I mentioned earlier. I wrapped the “Read Now” bookmarklet as a link for my purpose (as I’ll show in a bit). 

Instapaper too gives bookmarklets for readers – the “Instapaper Text” bookmarklet one is what I am interested in. For publishers Instapaper gives an URL that will add the page to the reader’s Instapaper queue. Unfortunately, there’s no similar URL to simply show the page in an Instapaper view without adding to queue. 

Pocket too gives bookmarklets and tools for publishers. However, both options only allow the page to be added to the Pocket queue, there’s no way to just get a Pocket view display of the page. 

So Readability and Instapaper are what I can use. I added the following text to each of my posts (I use the Atahualpa theme so it was just a matter of adding this text to the “Byline” section of each post and applying some formatting):

All this does is that it wraps the Javascript code from the Readability Read Now and the Instapaper Text bookmarklets within an a block like thus:

When a user clicks the text the JavaScript is executed as though they had clicked on the bookmarklet. (Hat tip to this page where I picked up the syntax from. I haven’t programmed with JavaScript so had to search around for how to invoke JavaScript from a link). 

Hope this helps someone! I have put links to both Instapaper and Readability because I prefer Instapaper but it is lousy with code blocks while Readability handles code blocks better. I feel Instapaper is better for text – it has more fonts, background colors, etc.

[Aside] What doesn’t seem like work?

From this essay by Paul Graham:

If something that seems like work to other people doesn’t seem like work to you, that’s something you’re well suited for.

That essay links to another longer essay about work and doing what you love. Have added it to my Instapaper queue.

Notes on AD Replication (contd.)

This post is a continuation to my previous one

How the AD Replication Model Works

(source)

Conflict Resolution

Previously I mentioned that conflict resolution in AD does not depend on timestamp. What is used instead of the “volatility” of changes. Here’s how it works in practice. 

Remember the replication metadata stored on each object? The ones you can view using repadmin /showobjmeta. I mentioned 5 metadata there – the Local USN, the Originating DSA, the Originating USN, the Originating Timestamp, and Version. Three of these metadata are used a conflict resolution stamp for every attribute:

  • Version, which as we know is updated each time the attribute is updated
  • Originating Timestamp, which is the timestamp from the DC where the update originated from
  • Originating DSA, which is not the DSA name or GUID as you’d expect from the repadmin output, but the invocationID of the DSA where the update originated from.

How is this stamp used? If there’s a conflict to an attribute – i.e. a change is made to an attribute on two separate DCs – first the Version is considered. Whichever update has the higher Version wins. Notice how the timestamp of the change doesn’t matter. Say WIN-DC01 had a change to an attribute twice (thus incrementing the Version twice) while WIN-DC02 had a change to the same attribute once, but at a later time, and both these changes reached WIN-DC03 together – the change from WIN-DC01 will win over the later change from WIN-DC02 because the number of changes were more there. 

If two conflicting changes have the same Version then the timestamp is considered. This has a one-second resolution, and so unless the conflict changes happened at the exact same second this is usually enough to resolve the conflict. 

However, if both Version and timestamp are unable to resolve the conflict, then the invocationID is considered. This is guaranteed to be different for each DC, and is random, so whichever change is from a DC with higher invocationID wins. 

Replication Metadata

The Knowledge Consistency Checker (KCC) (will be discussed in a later post) is the component that is responsible for maintaining the replication topology. It is maintains connection objects with the replication partners and stores this information, for each domain partition, in a multivalued attribute called repsFrom in the root of that domain partition. 

For example, here are the replication partners for WIN-DC02. Although not shown here, WIN-DC04 & WIn-DC05 are of a child domain. 

win-dc02-replications

Now consider the repsFrom attribute of the domain partition on WIN-DC02: 

And here’s the repsFrom from the Configuration partition:

Each entry starts from dwVersion and contains information like the number of failures, time of last successful sync, the DSA GUID, the database GUID, USNs, etc. Since only one DC is replicating with WIN-DC02 for the domain partition there’s only one value for that partition; while there are three DCs replicating for the Configuration partition and so there are three values for that partition. 

Each DC polls the DSAs (DCs) in this attribute for changes (that’s for the scheduled changes, not the ones where the source DC sends and update to all its partners and they poll for changes). If a DC is demoted – i.e. its NTDS settings object is deleted (i.e. the DSA is no longer valid) – the KCC will remove this DSA from the attribute. This prevents replication attempts to demoted DCs. (Prior to Windows 2003 though, and even now if this attribute is assigned a value, there used to be an attribute called replTopologyStayOfExecution. This value had a default of 14 days, and a maximum value of half the tombstone lifetime (the period for which deleted objects are retained). In the presence of this attribute – which existed by default in Window Server 2003 and prior, and can be set if required in later versions – if the KCC detects an invalid DSA, instead of removing it from the repsFrom attribute it will let it remain until such time the duration of the object being deleted exceeds replTopologyStayOfExecution). 

Atomicity

Atomicity is a term encountered in databases and operating systems (I first encountered it during my CS classes, specifically the OS course I think). An atomic operation can be thought of as an indivisible operation – meaning all events that take place during an atomic operation either take place together, or they don’t. (It comes from the idea that an atom was thought to be indivisible). With respect to databases, this is a guarantee that if a bunch of values are written in an atomic operation, either all the values are written or none of them. There’s no confusion that perhaps only a few values got committed while the rest were missed out. Even if one value didn’t get written, all others will not be written. That’s the guarantee! 

In the context of AD, updates are written to the AD database. And all attribute updates to a single object will be written atomically in a single transaction. However:

  • If the attributes are linked attributes (remember the previous post where there are attributes with forward and back links, for e.g. member and memberOf) the updates won’t be atomic – not too surprising, they are for different objects after all, and also usually the back link is generated by the system not sent as an update. 
  • Remember: the maximum number of values that can be written in a single transaction is 5000. 
  • To ensure that (nonlinked) attributes to an object are written atomically, updates to nonlinked attributes are prioritized over updates to linked attributes. This happens when a source DC packages all the updates into replication packets. The DC prioritizes nonlinked attributes over linked attributes. When it comes to writing the updates to the destination DC database though:
    • For linked attributes, because of parent-child relationships the objects might be written out of order to how the updates are received. This is to ensure that objects are created before any links are applied to that object.
    • When an object already exists on the destination DC, even though nonlined attributes are replicated first, they are not guaranteed to be written first to the database. Generally they are applied first, but it’s not guaranteed. (Note to self: I am not very clear about this point)
  • Remember: the number of values in a replication packet is approximately 100. If there are more than 100 values, again the nonlinked attributes are tried to put in one packet, while the linked attributes can span multiple packets. In such cases, when they are written on the destination DC database, all updates to a single object can require multiple transactions. (They are still guaranteed to be written in the same replication cycle). 
  • Note: Only originating updates must be applied in the same database transaction. Replicated updates can be applied in more than one database transaction.

 

Notes on AD Replication, Updates, Attributes, USN, High-Watermark Vector, Up-to-dateness Vector, Metadata, etc.

Reading a couple of AD Replication posts from TechNet/ MSDN. Below are my notes on them. 

Note

This is a super long post! It began as notes to the TechNet/ MSDN posts but it quickly grew into more. Lots of additional stuff and examples from my side/ other blog posts. Perhaps I should have split these into separate posts but I didn’t feel like it.

What is the AD Replication Model?

(source)

  • AD makes use of pull replication and not push. That is, each DC pulls in changes from other DCs (as opposed to the DC that has changes pushing these to targets). Reason for pulling instead of pushing is that only the destination DC knows what changes it needs. It is possible it has got some changes from another DC – remember AD is multimaster – so no point pushing all changes to a destination DC.
  • Initial replication to a new DC in the domain is via LDAP. Subsequent replications are via RPC.
  • All DCs do not replicate with all other DCs. They only replicate with a subset of DCs, as determined by the replication topology. (I am not sure, but I think this is only in the case of InterSite replication …). This is known as store and forward replication. DCs store updates they get from replication partners ams forward these to other DCs.
  • There are two sorts of replication – state-based and log-based.
    • State-based means DCs apply updates to their replicas (their copies of the partitions) as and when they arrive.
    • In log-based replication each DC maintains a log of the changes and sends this log to all other DCs. Once a destination DC receives this log it applies it to its replica and becomes up-to-date. This also means the destination DC will receive a list of all changes (from the source DC perspective), not just the changes it wants. 
    • Active Directory (actually called Active Directory Domain Services since Server 2008) uses state-based replication. 
    • Each DC maintains a list of objects by last modification. This way it can easily examine the last few objects in the list to identify how much of the replication source’s changes have already been processed. 

Linked & Multivalued attributes

Before the next post it’s worth going into Linked attributes and Multivalued attributes.

  • Linked attributes are pairs of attributes. AD calculates the value of one of the attributes (called the “back link”) based on the value of the other attribute (called the “forward link”). For example: group membership. Every user object has an attribute called memberOf. You can’t see this by default in ADUC in the “Attribute Editor” tab. But if you click “Filter” and select to show “Backlinks” too then you can view this attribute. backlinks The memberOf attribute is an example of a back link. And it’s generated by AD, which is why it’s a read-only attribute. It’s counterpart, the forward link, is the member attribute of a group object. Linked attributes are linked together in the schema. In the schema definitions of those attributes – which are instances of an attributeSchema class – there is a linkID attribute that defines the back and forward links. The attribute with an even linkID is the forward link (the write-able attribute that we can set) while the attribute with a linkID that’s one more than the linkID of the forward link is the back link (the read-only, AD generated attribute). For instance, using ADSI Edit I found the CN=Member,CN=Schema,CN=Configuration,DC=mydomain object in the Schema partition. This is the schema definition of the member attribute. Being a forward link it has an even linkID. linkid Its counterpart can be easily found using ldp.exe. Search the Schema partition for all objects with linkID of 3. linkid-2 The result is CN=Is-Member-Of-DL,CN=Schema,CN=Configuration,DC=mydomain, which is the attributeSchema object defining the memberOf attribute (notice the lDAPDisplayName attribute of this object in the above screenshot).
    • It’s worth taking a moment to understand the above example. When one uses ADUC to change a user/ group’s group membership, all one does is go to the user/ group object, the “Member Of” tab, and add a group. This gives the impression that the actual action happens in the user/ group object. But as we see above that’s not the case. Although ADUC may give such an impression, when we add a user/ group to a group, it is the group that gets modified with the new members and this reflects in the user/ group object that we see. ADUC makes the change on the user/ group, but the DSA makes the real change behind the scenes. The cause and effect is backwards here …
    • Another obvious but not so obvious point is that when you add a user/ group to a group, it is the group’s whenChanged attribute that gets changed and it is the change to the group that is replicated throughout the domain. The user object has no change and nothing regarding it is replicated. Obvious again, because the change really happens on the group and the effect we see on the user/ group is what AD calculates and shows us. 
  • Multivalued attributes, as the name suggests, are attributes that can hold multiple values. Again, a good example would be group membership. The member and memberOf attributes mentioned above are also multivalued attributes. Obvious, because a group can contain multiple members, and a user/ group can be a member of multiple groups. Multivalued attributes can be identified by an attribute isSingleValued in their attributeSchema definition. If this value is TRUE it’s a single valued attribute, else its a multivalued attribute.

How the AD Replication Model Works

(source)

A must read post if you are interested in the details! 

On replication

  • AD objects have attributes. AD replicates data at the attribute level – i.e. only changes to the attributes are replicated, not the entire object itself. 
    • Attributes that cannot be changed (for e.g. back links, administrative attributes) are never replicated. 
  • Updates from the same directory partition are replicated as a unit to the destination DC – i.e. over the same network connection – to optimize network traffic.
  • Updates are always specific a single directory partition. 
  • Replication of data between DCs consist of replication cycles (i.e. the DCs keep replicating until all the data has replicated). 
    • There’s no limit to the number of values that can be transferred in a replication cycle. 
    • The transfer of data happens via replication packets. Approximately 100 values can be transferred per packet. 
      • Replication packets encode the data using ASN.1 (Abstract Syntax Notation One) BER (Basic Encoding Rules). Check out Wikipedia and this post for more information on ASN.1 BER (note: ASN.1 defines the syntax; BER defines the encoding. There are many other encodings such as CER, DER, XER but BER is what LDAP uses). See this blog post for an example LDAP message (scroll down to the end). 
    • Once a replication cycle has begun, all updates on the source DC – including updates that occur during the replication cycle – are send to the destination DC. 
    • If an attribute changes multiple times between replication cycles only the latest value is replicated. 
    • As soon as an attribute is written to the AD database of a DC, it is available for replication. 
  • The number of values that can be written in a single database transaction (i.e. the destination DC has all the updates and needs to commit them to the AD database) is 5000. 
    • All updates to a single object are guaranteed to be written as a single database transaction. This way the state of an object is always consistent with the DC and there are no partially updated objects. 
      • However, for updates to a large number of values to multivalued attributes (e.g. the member attribute from the examples above) the update may not happen in the same transaction. In such cases the values are still guaranteed to be written in the same replication cycle (i.e. all updates from the source DC to that object will be written before the DC performs another replication cycle). (This exception for multivalued attributes is from Server 2003 and above domains, I think).
    • Prior to Windows Server 2003, when a multivalued attribute had a change the entire attribute was replicated (i.e. the smallest change that could be replicated was an attribute). 
      • For example, if a multivalued attribute such as member (referring to the members of a group) had 300 entries and now an additional entry was added, all 301 entries were replicated rather than just the new entry. 
      • So if a group had 5000 members and you added an extra member, 5001 entries were replicated as updates to its member attribute. Since that is larger than the number of values that can be committed in a single transaction, it would fail. Hence the maximum number of members in a group in a Windows 2000 forest functional level domain was 5000. 
    • Starting from Windows Server 2003, when a multivalued attribute has a change only the changed value is replicated (i.e. the smallest change that can be replicated is a value). Thus groups in Windows Server 2003 or Windows Server 2003 interim functional level domains don’t have the limitation of 5000 members. 
      • This feature is known as Linked Value Replication (LVR). I mentioned this in an earlier post of mine in the context of having multiple domains vs a single domain. Not only does LVR remove limitations such as the above, it also leads to efficient replication by reducing the network traffic.
      • When a Windows 2000 forest level domain is raised to Windows 2003/ 2003 interim forest level, existing multivalued attributes are not affected in any way. (If they were, there’d be huge network traffic as these changes propagate through the domain). Only when the multivalued attribute is modified does it convert to a replicate as single values. 
  • Earlier I mentioned how all attributes are actually objects of the class attributeSchema.  The AD schema is what defines attributes and the relations between them.
    • Before replication between two DCs can happen, the replication system (DRS – see below) checks whether the schema of both DCs match. If the schemas do not match, replication is rescheduled until the schemas match (makes sense – if replication happens between DCs with different schemas, the attributes may have different meanings and relationships, and replication could mess things up). 
    • If the forest schema is changed, schema replication takes precedence over all other replication. In fact, all other replication is blocked until the schema replicates. 

Architecture

  • On each DC, replication operates within the Directory System Agent (DSA) component – ntdsa.dll
    • DSA is the interface through which clients and other DCs gain access to the AD database of that DC. 
    • DSA is also the LDAP server. Directory-aware applications (basically, LDAP-aware applications) access DSA via the LDAP protocol or LDAP C API, through the wldap32.dll component  (am not very clear about this bit). LDAPv3 is used. 
  • A sub-component of the DSA is DRS (Directory Replication System) using the DRS RPC protocol (Microsoft’s specification of this protocol can be found at this MSDN link). Client & server components of the DRS interact with each other to transfer and apply updates between DCs. 
    • Updates happen in two ways: via RPC or SMTP (the latter is only for non-domain updates, which I infer to mean Schema or Configuration partition updates) 
    • Domain joined computers have a component called ntdsapi.dll which lets them communicate with DCs over RPC.
      • This is what domain joined computers use to communicate with the DC.
      • This is also what tools such as repadmin.exe or AD Sites and Services (dssites.msc) use to communicate with the DC (even if these tools are running on the DC). 
    • DCs additionally have a private version of ntdsapi.dll which lets them replicate with other DCs over RPC. (I infer the word “private” here to mean it’s meant only for DCs).
      • As mentioned earlier, DC to DC replication can also bypass RPC and use SMTP instead. Then the ntdsapi.dll component is not used. Other components, such as ISMServ.exe and the CDO library are used in that case. Remember: this is only for non-domain updates. 
    • The RPC interface used by DRS is called drsuapi. It provides the functions (also see this link) for replication and management. Check this section of the TechNet post for an architecture diagram and description of the components. 
      • The TechNet post also mentions DRS.idl. This is a file that contains the interface and type library definitions. 
  • The DSA also has a sub-component for performing low-level operations on the AD database. 
  • The DSA also has a sub-component that provides an API for applications to access the AD database. 
    • The AD database is called ntds.dit (DIT == Directory Information Tree; see this link for more info). The database engine is Extensible Storage Engine (ESE; esent.dll)

Characteristics of AD replication

  • Multimaster
    • Changes can be made to any DCs (as long as they are authoritative for the objects).
  • Pull
    • When an update occurs on a DC it notifies its partners. The partners then requests (pulls) changes from the source DC. 
  • Store-and-forward
    • When a DC receives a change from its partners, it stores the change and in-turn forwards on to others (i.e. informs others so they can issue pull requests from itself). Thus, the DC where a change is made does not have to update every other DCs in the domain. 
    • Because of the store-and-forward mechanism replication must be thought of as sequentially moving through the domain. 
  • State-based
    • Each DC has metadata to know the “state” of its replicas. This state is compared with that of its partner to identify the changes required. (This is in contrast to log-based replication where each DC keeps a log of changes it made and sends that log to its partners so they can replay the log and update themselves). 
    • This makes uses of metadata such as Update Sequence Number (USN), Up-to-dateness vector, and High-watermark vector. Synchronized time is not primarily used or required to keep track of this (it is used, but in addition to other mechanisms). 

Updates

LDAP/ AD supports the following four types of update requests:

  • Add an object to the directory.
  • Delete an object from the directory. 
  • Modify (add, delete, remove) attribute values of an existing object in the directory. 
  • Move an object by changing the name or parent of the object. 

Each of the above update requests generates a separate write transaction (because remember, all updates to a single object happen in a single transaction). Updates are an all-or-nothing event. If multiple attributes of an object are updated, and even one of those updates fail when writing to the database, all attributes fail and are not updated. 

Once an update request is committed to the database it is called an originating update. When a DC receives an originating update, writes it to its database, and sends out updates to other DCs these are called replication updates. There is no one-to-one relation between originating updates and replication updates. For instance, a DC could receive multiple originating updates to an object – possible even from different DCs – and then send a replication update with the new state of that object. (Remember: AD is state-based, not log-based. It is the state that matters, not the steps taken to reach that state). 

Adding request:

  • Creates a new object with a unique objectGUID attribute. Values of all replicated attributes are set to Version = 1. (The Version attribute will be discussed later, below). 
  • The add request fails if the parent object does not exist or the DC does not contain a writeable replica of the parent object’s directory partition. 

Modify request:

  • If an attribute is deleted then its value is set with NULL and Version is incremented. 
  • If values are to be added/ removed from an attribute, it is done so and the attribute Version is incremented. The Modify request compares the existing and new values. If there are no changes then nothing happens – the request is ignored, the Version does not change. 

Move request:

  • This is a special case of the Modify request in that only the name attribute is changed. 

Delete request:

  • The isDeleted attribute is set to TRUE. This marks the object as tombstoned (i.e. deleted but not removed from AD). 
  • The DN of the object is set to a value that cannot be set by an LDAP application (i.e. an impossible value). 
  • Except a few important attributes, the rest are stripped from the object (“stripped”, not removed as one would do above). 
  • The object itself is moved to the Deleted Objects container. 

Certain objects are protected from deletion:

  • Cross-references (class crossRef) and RID Object for the DC. Any delete request for these objects are rejected. And any originating update deletion of these objects is rejected, and all attributes of the object are updated (Version incremented) so the object replicates to other DCs and is reanimated wherever it was deleted.  
    • Cross-references are present in the Configuration partition: CN=Partitions,CN=Configuration,DC=(domain)
    • RID Objects are present under each DC at CN=RID Set,CN=(DC),OU=Domain Controllers,DC=(domain)
  • The NTDS Settings (class nTDSDSA) object.
    • This object represents the DSA (Directory System Agent). It is present in the Configuration partition under the site and DC: CN=NTDS Settings,CN=(DC),CN=Servers,CN=(SITE),CN=Sites,CN=Configuration,DC=(domain)
    • Remember, the DSA is the LDAP server within the DC. The DSA’s GUID is what we see when DCs replicate or in their DNS names (DSAGUID._msdcs.domain is a CNAME to the DC name).
    • The objectGUID of the DC object is not the same as the objectGUID of its DSA object. The latter is what matters. When a DC computer object is opened in ADUC, it is possible to view the corresponding DSA object by clicking on the “NTDS Settings” button in the first tab. 
    • Trivia: to find all the DCs in a forest one can query for objects of class nTDSDSA. Below query, for instance, finds all objects of that class and returns their DN and objectGUID

      nTDSDSA

    • When a DC is demoted, its DSA object is deleted but not really deleted. It is protected and disabled from receiving replication requests. Thus, a query such as the above will return DSA objects of DCs that may no longer exist in the forest.

      A better way to structure the above query then would be to also filter out objects whose isDeleted attribute is not TRUE.

      nTDSDSA-2

Conflicts

AD does not depend on timestamps to resolve conflicts. In case of conflicts the “volatility” of changes is what primarily matters. That is, if an attribute was updated twice on one DC and thrice on another DC, even if the timestamps of changes from the first DC are later than the second DC, the change from the second DC will win because the attribute was updated thrice there. More on this in my next post

Database GUID

I mentioned above that each DSA has its own objectGUID. This GUID is created with the server is promoted to a DC and deleted (sort of) when the DC is demoted. Thus the GUID is present for the lifetime of the DC and doesn’t change even if the DC name changes. 

The AD database (ntds.dit) on each DC has its own GUID. This GUID is stored in the invocationId attribute of the DSA. Unlike the DSA GUID, the database GUID can change. This happens when (1) the DC is demoted and re-promoted (so the database changes), (2) when an application NC is added/ removed to the DC, (3) when the DC/ database is restored from a backup, or (4) (only in Server 2012 and above) whenever a DC running in a VM is snapshotted or copied.

The invocationId attribute can be viewed via ldp.exe as above, or in the “NTDS Settings” of the DC computer object in AD Users & Computers, or in the “NTDS Settings” in AD Sites & Services. It can also be viewed in the output of repadmin.exe.

The invocationID is a part of replication requests (more later). When the invocationID changes other know that they have to replicate changes to this DC so it can catch up. The first DC in a domain will have its invocationID and objectGUID as same (until the invocationID changes). All other DCs will have different values for these both. 

Update Sequence Numbers (USNs)

(I was planning on covering this separately as part of my AD Troubleshooting WorkshopPLUS posts, but USNs are mentioned in this TechNet post so I might as well cover them now). 

USNs are 64-bit counters maintained on each DC. The number is different for each DC and is stored in the highestCommittedUsn attribute of the rootDSE object.

rootDSE is an imaginary object. It is the root of the directory namespace. Under it we have the various domain partitions, configuration partition, schema partition, and application partitions. rootDSE can be queried to find the various partitions known to the DC. It can be viewed through ADSI Edit or ldp.exe (the latter makes it easier to view all the attributes together).

rootDSE

As can be seen in the screenshot above, one of the attributes is highestCommittedUSN.

Each time a DC commits an update (originating update/ replication update), it updates the highestCommittedUSN attribute. Thus the USN is associated with each successful transaction to the AD database. Remember: all updates to a single object are written in a single transaction, so a USN is essentially associated with each successful update to an object – be it a single attribute, or multiple updates to the same/ multiple attributes (but all written in the same transaction). 

USNs are update incrementally for each write transaction. So, for example, when I took the above screenshot the USN for WIN-DC02 was 77102. When I am writing this paragraph (the next day) the USN is 77230. This means between yesterday and today WIN-DC02 has written 128 transactions (77230-77102) to its database.  

Every attribute of an object has a USN associated with it. This is a part of the replication metadata for that object, and can be viewed through the repadmin.exe command. For instance, using the /showobjmeta switch (note that here we specify DC first and then object):

Notice how the attribute USNs vary between the two DCs. Also notice the metadata stored – the Local USN, the originating DSA, the Originating DSA's USN, the timestamp of the update, and Version. If you are running the above command it is best to set the command prompt window width to 160, else the output doesn’t make much sense. I will talk about Local USN and Originating USN in a moment. 

Another switch is /showmeta (here we specify the object first and then the DC):

Both switches seem to produce the same output. The important items are the Local USN and Originating USN, and the Version

  • Version starts from 1 and is incremented for each update. Version will be same on all DCs – when an DC commits an update request (originating update or replication update) it will increment the Version. Since all attributes start with the same Version = 1, the current value will be the same on all DCs. 
  • Local USN is the USN value for that attribute on the DC we are looking at. It is the value of the highestCommittedUSN for the transaction that committed the update to this attribute/ set of attributes
  • Originating USN is the USN for that attribute on the DSA where the originating update was sent from.

For instance: the attribute description. WIN-DC01 has local USN 46335, WIN-DC02 has 19340, and WIN-DC03 has 8371. The latter two DCs got their update for this attribute from WIN-DC01 so they show the originating DSA as WIN-DC01 and the Originating USN as 46335.

Every object has an attribute called uSNChanged. This is the highest Local USN among all the attributes of that object. (What this means is that from the USN of an object we can see which of its attributes have the same local USN and so easily determine the attributes changed last). Unlike the attribute USNs which are metadata, uSNChanged is an object attribute and thus can be viewed via ADSI Edit or ADUC. From the command line, repadmin.exe with the /showattr switch can be used to view all attributes of an object (this switch takes the DC name first, then the object). 

Above output tells me that on WIN-DC01 the uSNChanged for this object is 46335.  From the earlier /showobjmeta output I know that only the description attribute has that local USN, so now I know this was the last attribute changed for this object. 

So, to recap: DCs have a USN counter stored in highestCommitedUSN attribute. This counter is updated for each successful write transaction to the database. Since writes happen per object, it means this USN is updated every time an object is updated in the database. Each object attribute has its own USN – this too is per DC, but is not an attribute, it is metadata. Finally, each object has a uSNChanged attribute which is simply the highest attribute USN of that object. This too is per DC as the attribute USN is per DC. The DC’s highestCommittedUSN attribute and an object’s uSNChanged attribute are related thus: 

  • When an attribute update is committed to the database the DC’s highestCommittedUSN is incremented by 1.
  • The Local USN of the attribute/ attributes is set to this new highestCommittedUSN
  • This in turns updates the object’s uSNChanged to be the new highestCommittedUSN (because that is the highest attribute Local USN now)

Thus, the highestCommittedUSN is the highest uSNChanged attribute among all the replicas held by the DC.  

Here’s an example. First I note the DC’s highestCommitedUSN (I have removed the other attributes from the output). 

Then I note an object’s uSNChanged:

Now I connect via ADUC and change the description field. Note the new USN. 

And note the DC’s USN:

It has increased, but due to the other changes happening on the DC it is about 10 higher than the uSNChanged of the object I updated. I searched the other replicas on my DC for changes but couldn’t find anything. (I used a filter like (uSNChanged>=125979) in ldp.exe and searched every replica but couldn’t find any other changes. Maybe I missed some replica – dunno!) This behavior is observed by others too. (From the previous linked forum post I came across this blog post. Good read).  

Finally, I must point out that even though I said the attribute metadata (Local USN, Originating USN, Originating DSA, Originating Time/ Date, Version) are not stored as attributes, that is not entirely correct. Each object has an attribute called replPropertyMetaData. This is an hexadecimal value that contains the metadata stored as a binary value. In fact, if we right click on an object in ldp.exe it is possible to view the replication metadata because ldp.exe will read the above attribute and output its contents in a readable form. 

ldp-metadata

Bear in mind this attribute is not replicated. It is an attribute that cannot be administratively changed, so can never be updated by end-users and doesn’t need replicating. It is only calculated and maintained by each DC. uSNChanged is a similar attribute – not replicated, only administratively changed. 

Note to self: I need to investigate further but it looks like uSNChanged cannot be viewed by ldp.exe for all objects. For instance, ldp.exe shows this attribute for one user object in my domain but not for the other! I think that’s because the attribute is generated by the AD tools when we view it. repadmin.exe and the GUI tools show it always. Similarly, the attribute level metadata attribute too cannot be viewed by ldp.exe for all objects. For some objects it gives an error that the replPropertyMetaData attribute cannot be found and so cannot show the replication metadata. This could also be why there was a gap between the uSNChanged and highestCommittedUSN above. 

High-watermark Vector (HWMV) & Up-to-dateness Vector (UTDV)

To recap once more, what we know so far is:

  • Every attribute has replication metadata that contains its Local USN, the originating DSA's USN, the originating DSA, the timestamp, and Version
  • The highest Local USN of an attribute is stored as the object’s uSNChanged attribute. 
  • The highest uSNChanged attribute among all objects on all replicas held by the DSA is stored as its highestCommittedUSN attribute in the rootDSA
  • All these USN counters are local to the DC (except for the Originating USN which is local to that the originating DSA). 

How can DCs replicate with each other using the above info? Two things are used for this: High-watermark vector and Up-to-dateness vector. 

The HWMV is a table maintained on each DC.

  • There is one table for each directory partition the DC holds replicas for (so at minimum there will be three HWMV tables – for the domain partition, the Schema partition, and the Configuration partition). 
  • The HWMV table contains the highest USN of the updates the DC has received from each of its replication partners for that replica. Note: the table contains the USN that’s local to the replication partner.
  • Thus the HWMV can be thought of as a table contain the highest uSNChanged value for each directory partition of its replication partners when they last replicated.

Got it? Now whenever a DC needs to replicate with another DC, all it needs to do is ask it for each of the changes since the last uSNChanged value the destination is aware of! Neat, right! The process looks roughly like this:

  1. The originating DC will have some changes. These changes will cause uSNChanged values of the affected objects to change. The highestCommittedUSN of the DC too changes. All these are local changes.
  2. Now the DC will inform all its replication partners that there are changes. 
  3. Each of these replication partners will send their HWMV for the partition in question.
  4. From the received HWMV table, the originating DC can work what changes need to be sent to the destination DC. 
    1. The originating DC now knows what USNs it needs to look for. From this it knows what objects, and in turn which of their attributes, were actually changed. So it sends just these changed attributes to the destination DC. 

To make the process of replication faster, all DCs have one more table. That is the Up-to-dateness Vector (UTDV).

  • Like the HWMV, this too is for every replica the DC holds. 
  • Unlike the HWMV this contains the highest USN of the updates the DC has received from every other DC in the domain/ forest for that replica. 
  • The table also contains the timestamp of last successful replication with each of those DCs for that replica.

The UTDV table has a different purpose to the HWMV. This table is sent by the destination DC to the originating DC when it requests for changes – i.e. in step 3 above.

When the originating DC gets the UTDV table for its destination DC, it can look at the table and note the destination DC’s highest received USNs for that partition from other DCs. Maybe the destination DC has asked for changes from USN number x upwards (the USN number being of the originating DC). But these changes were already received by the destination DC from another DC, under USN number y and upwards (the USN number being of that other DC). The destination DC does not know that the changes requested by USN x and upwards are already with it from another DC, but by looking at the UTDV table the originating DC can see that USNs y and above already contain the changes the destination DC is requesting, so it can filter out those updates when sending. (This feature is called “propagation dampening”). 

  • In practice, once the originating DC compiles a list of USNs that need to be sent to the destination DC – at step 4 above – it goes through its replication metadata to check each attribute and the originating DSA and originating USN associated with that attribute.
  • The originating DC then looks at the UTDV table of the destination DC, specifically at the entry for the DC that had sent it an update for the changed attribute. (This DC could be same as the originating DC). From the table it can see what USNs from this DC are already present at the destination DC. If the USN value in the table is higher than the USN value of the attribute, it means the destination DC already has the change with it. So the originating DC can filter out these and similar attributes from sending to the destination DC.  

Thus the UTDV table works along with the HWMV table to speed up replication (and also avoid loops wherein one DC updates another DC who updates the first DC and thus they keep looping). And that is how replications happen behind the scenes! 

Once a destination DC updates itself from an originating DC – i.e. the replication cycle completes – the source DC sends its UTDV table to the destination DC. The destination DC then updates its UTDV table with the info from the received UTDV table. Each entry in the received table is compared with the one it has and one of the following happens:

  • If the received table has an entry that the destination DC’s UTDV table does not have – meaning there’s another DC for this replica that it isn’t aware of, this DC has replicated successfully with the originating DC and so all the info it has is now also present with the destination DC, and so it is as good as saying this new DC has replicated with the destination DC and we are aware of it the same way the originating DC is aware – so a new entry is added to the destination DC’s UTDV table with the name of this unknown DC and the corresponding info from the received UTDV table. 
  • If the received table has an entry that the destination DC’s UTDV table already has, and its USN value is higher than what the destination DC’s table notes – meaning whatever changes this known DC had for this partition has already replicated with the originating DC and thus the destination DC – and so its entry in the UTDV can actually be updated, the UTDV table for that server is updated with the value from the received UTDV table.  

The UTDV table also records timestamps along with the USN value. This way DCs can quickly identify other DCs that are not replicating. These timestamps record the time the DC last replicated with the other DC – either directly or indirectly. 

Both HWMV and UTDV tables also include the invocationID (the database GUID) of the DCs. Thus, if a DC’s database is restored and its invocationID changes, other DCs can take this into account and replicate any changes they might have already replicated in the past.  

From the book I am reading side-by-side (excellent book, highly recommended!) I learnt that apart from the HWMV and UTDV tables and the naming context it wants to replicate, the destination DC also sends two other pieces of info to originating DC. These are: (1) the maximum number of object updates the destination DC wishes to receive during that replication cycle, and (2) the maximum number of values the destination DC wishes to receive during that replication cycle. I am not entirely clear what these two do. Once a replication cycle begins all object updates and values are sent to the destination DC, so the two pieces above seems to be about whether all the updates are sent in one replication packet or whether they are split and sent in multiple packets. The maximum number of values in a single packet is about 100, so I guess these two numbers are useful if you can only accept less than 100 values per packet – possibly due to network constraints. 

More later …

Unfortunately I have to take a break with this post here. I am about halfway down the TechNet post but I am pressed for time at the moment so rather than keep delaying this post I am going to publish it now and continue with the rest in another post (hopefully I get around to writing it!). As a note to myself, I am currently at the Active Directory Data Updates section, in the sub-section called “Multimaster Conflict Resolution Policy”.

[Aside] Some quotes

On passion

I think it’s not exactly true to say that if you do something you are passionate about, the money will come. But, if you do almost anything really diligently then the money will come. And it takes passion to have that kind of diligence. And … if you almost anything really diligently you will figure out what parts of it like and what you don’t and that will help inform your choices (when it comes to choosing a job or doing another thing). 

– Isaac Schlueter (from this podcast)

On money

Money is not the root of all evil. The love for money is the root of all evil. 

– Apparently this is the original version but somewhere along the line became misquoted. I came across this via another podcast (I think Jonathan Strickland said the correct version on the podcast)