Network Working Group                                   Paul Mockapetris
Request for Comments: 973                                            ISI
                                                            January 1986

                 Domain System Changes and Observations


STATUS OF THIS MEMO

   This RFC documents updates to Domain Name System specifications
   RFC-882 [1] and RFC-883 [2], suggests some operational guidelines,
   and discusses some experiences and problem areas in the present
   system.  Distribution of this memo is unlimited.

   This document includes all changes to the Domain System through
   January, 1986.  Change notices and additional discussion are
   available online in file [USC-ISIB.ARPA]<DOMAIN>DOMAIN.CHANGES.

OVERVIEW

   This memo is divided into four major sections:

      "UPDATES" which discusses changes to the domain specification
      which are in widespread use and should be regarded as being part
      of the specification.

      "OPERATION GUIDELINES" which suggests rules-of-thumb for using the
      domain system and configuring your database which are appropriate
      in most cases, but which may have rare exceptions.

      "EXPERIENCES" which discusses some unusual situations and common
      bugs which are encountered in the present system, and should be
      helpful in problem determination and tuning.

      "PROBLEM AREAS" which discusses some shortcomings in the present
      system which may be addressed in future versions.

UPDATES

   This section discusses changes to the specification which are final,
   and should be incorporated in all domain system software.

   TTL timeouts too small

      The 16 bit TTL field in RRs could not represent a large enough
      time interval.  The 16 bit field, using seconds for units, has a
      maximum period of approximately 18 hours.

      All time values, including all TTLs and the MINIMUM field of the
      SOA RR, are expanded to 32 bits.

   CLASS changes

      Class 2, originally reserved for CSNET, is obsolete.  Class 3 has
      been assigned for use by CHAOS.

   CNAME usage

      The specification allows CNAME RRs to exist with other RRs at the
      same node.  This creates difficulties since the other RRs stored
      with the CNAME at the alias might not agree with the RRs stored at
      the primary name.

      If a node has a CNAME RR, it should have no other RRs.

   * semantics

      The use of * to represent a single label wildcard, along with the
      possibility of multiple * labels, led to difficult server
      implementations and complicated search algorithms.  There were
      also questions regarding whether a * based specification could
      refer to names that were not contained in the zone which had the *
      specification.

      While we might want the "inheritability" for some cases, it leads
      to implementation difficulties.  The first of these is that
      whenever we can't find a RR in a particular zone, we have to
      search all parent zones to look for a suitable * result.
      (Alternatively we could develop some automatic method for insuring
      consistency or insist on careful duplication of inherited data.)
      We also must deal with conflicts, i.e. what if a subdomain doesn't
      want to inherit defaults.

      Given these difficulties, the solution is to insist that
      delegation of authority cancels the * defaults.  This is quite
      simple to implement; all you need to do is to check for delegation
      before looking for * RRs.

      A second difficulty is the restriction that * match a single
      label.  Thus if a name server is looking for RRs for the name
      A.B.C.D.E.F, it must check for *.B.C.D.E.F, *.*.C.D.E.F,
      *.*.*.D.E.F, etc.  This check must also be careful of zone
      boundaries and multiplies the effort to handle a query.

      The solution adopted is to allow a single * label in the leftmost
      part of a name stored in a zone, and to allow this label to match

      any number of unknown labels or a single known label in the query
      name.  However, the * match is only taken for parts of the tree
      which are neither delegated or explicitly represented.

      The algorithm for performing the search in a tree structured
      database has the following steps:

      1) Descend in the tree matching labels from right to left.  If a
      delegation is found return that;  if the specified node is found
      go to step 2, if the tree ends go to step 3.

      2) Look for RRs that answer the query.  If any are found, return
      them as the answer.  If none are found, look for answers in a *
      node which has the same name as the query name except for the
      rightmost label.  (e.g. if you can't find an answer at F.ISI.ARPA,
      look for a RR at *.ISI.ARPA)

      3) The search for a desired name has failed; look for a node whose
      name is * plus however much matched.  Look for answers there.
      (e.g. If you are looking for X.Y.ISI.ARPA and the tree ends at
      ISI.ARPA, look at *.ISI.ARPA.  The same thing holds for
      Y.ISI.ARPA, or any name of the form <anything>.Z.ISI.ARPA, where Z
      is a label that doesn't exist under ISI.ARPA)

      Note that this interpretation means that * matches names that are
      not in the tree, no matter how much of the tree is missing, and
      also matches one level's worth of known tree.

   AA semantics

      When a name server is responding to a query for a particular name
      and finds a CNAME, it may optionally restart the search at the
      canonical name.  If the server uses the restart feature, the
      answer section of the returned query contains one (or more)
      CNAMEs, possibly followed by answers for the primary name.  The
      canonical name will usually be in the same zone as the alias, but
      this need not be the case.  If the server is authoritative for one
      of the names but not both, it is not clear whether the AA bit
      should be set.

      The solution adopted is to make the AA refer to the original query
      name.

   Master file format

      The present specification uses a somewhat awkward method for
      representing domain names in master files.

      The change adopted is that all domain names in this file will be
      represented as either absolute or relative.  An absolute domain
      name ends with a ".".  A free standing "." is assumed to refer to
      the root.  A relative domain name doesn't end with a dot, and is
      assumed to be relative to the current origin.

   SERIAL number size

      If the master file changes rapidly, an infrequently updated copy
      may miss the wrapping of the sequence number in the SERIAL field
      of the SOA, or misinterpret the number of updates that have taken
      place.

      The SERIAL field is increased to 32 bits.

   MD and MF replaced by MX

      The original specification uses MD and MF RRs for mail agent
      binding.  The problem is that a mailer making a MAILA query, which
      asks for both types, can't use the cache since the cache might
      have the results for a MD or MF query.  That is, the presence of
      one of these types of information in the cache doesn't imply
      anything about the other type.  The result was that either mailers
      would have to always consult authoritative servers or try to use
      partial information; neither of these is really acceptable.

      The change is to replace MD and MF with a new type of RR called MX
      which conveys similar information in a single RR type.  MX has
      been assigned a type code of 15 decimal.  The format of the MX RR
      is a 16 bit preference value followed by a domain name.  A node
      may have multiple MX RRs, and multiple MX RRs with the same
      preference value are allowed at a given node.

      The preference values denote the relative preference that the mail
      destination places on the mail agents, with lower values being
      "better".  A mailer is expected to at least try the mail agent(s)
      with the lowest preference value.  The significance of particular
      preference values, the units of preference, and the linearity of
      preference values are not defined but left open; preference values
      should only be used to establish relative rankings.

      For example, the current RRs:

                       MAIL-ORG   MD    HOST1    
                                  MD    HOST2    
                                  MF    HOST3    

      might be replaced by:

                       MAIL-ORG   MX    10 HOST1 
                                  MX    10 HOST2 
                                  MX    20 HOST3 

      The values 10 and 20 have no significance other than 10<20.  A
      detailed discussion of the use of MX is the subject of [3].

   Zone transfer

      The original specification states that zone transfers take place
      in breadth first order.  The intent was to make the transfer
      easier for the accepting name server to handle.  This now doesn't
      work out to be very helpful, and is a severe pain for implementers
      using various hashing algorithms.  The new rule is that you can
      transmit the records in any order you choose, so long as the SOA
      node of the zone is transmitted first and last, and no other
      duplication occurs.

   IN-ADDR domain renamed

      The name of the IN-ADDR domain is now IN-ADDR.ARPA.  This change
      was made because many felt that the use of a top-level name was
      inappropriate to network-specific information.

OPERATIONAL GUIDELINES

   This section suggests rules-of-thumb for using the domain system and
   configuring your database which are appropriate in most cases, but
   which may have rare exceptions.

   Zone delegation

      When a domain wishes to become independent from its parent, the
      RRs which mark the delegation in the parent and child zones should
      be carefully synchronized to minimize the possibility that
      resolvers become confused.

      For example, suppose that we wish to create a new zone called
      ISI.EDU under an existing EDU zone, and that the servers for the
      child zone are X.ISI.EDU and Y.GOV.

      We might add the following to the parent zone:

       ISI.EDU.      10000 NS  X.ISI.EDU.              
                     10000 NS  Y.GOV.                  
       X.ISI.EDU.    10000 A   <address of X.ISI.EDU.> 
       Y.GOV.        10000 A   <address of Y.GOV.>     

      and the following to the child zone:

       ISI.EDU.      10000 NS  X.ISI.EDU.              
                     10000 NS  Y.GOV.                  
                     50000 SOA <SOA information>       
       X.ISI.EDU.    10000 A   <address of X.ISI.EDU.> 
       Y.GOV.        10000 A   <address of Y.GOV.>     

      Note the following:

         In both cases, the A RR for Y.GOV is included, even though
         Y.GOV isn't in the EDU or ISI.EDU domains.  This RR isn't
         authoritative, but is included to guarantee that the address of
         Y.GOV is passed with delegations to it.  Strictly speaking this
         RR need not be in either zone, but its presence is recommended.
         The X.ISI.EDU A RR is absolutely essential.  The only time that
         a server should use the glue RRs is when it is returning the NS
         RRs and doesn't otherwise have the address of the server.  For
         example, if the parent server also was authoritative for GOV,
         the glue RR would typically not be consulted.  However, it is
         still a good idea for it to be present, so that the zone is
         self-sufficient.

         The child zone and the parent zone have identical NS RRs for
         the ISI.EDU domain.  This guarantees that no matter which
         server is asked about the ISI.EDU domain, the same set of name
         servers is returned.

         The child zone and the parent zone have A RRs for the name
         servers in the NS RRs that delegate the ISI.EDU domain.  This
         guarantees that in addition to knowing the name servers for the
         ISI.EDU domain, the addresses of the servers are known as well.

         The TTLs for the NS RRs that delegate the ISI.EDU domain and
         the A RRs that represent the addresses of the name servers are
         all the same.  This guarantees that all of these RRs will
         timeout simultaneously.  In this example, the value 10000 has
         no special significance, but the coincidence of the TTLs is
         significant.

      These guidelines haven't changed any of the flexibility of the
      system; the name of a name server and the domains it serves are
      still independent.

      It might also be the case that the organization called ISI wanted
      to take over management of the IN-ADDR domain for an internal
      network, say 128.99.0.0.  In this case, we would have additions to
      the parent zone, say IN-ADDR.ARPA.

      We might add the following to the parent zone:

       99.128.IN-ADDR.ARPA. 2000 NS  Q.ISI.EDU.               
                            2000 NS  XX.MIT.EDU.              
       Q.ISI.EDU.           2000 A   <address of Q.ISI.EDU.>  
       XX.MIT.EDU.          2000 A   <address of XX.MIT.EDU.> 

      and the following to the child zone:

       99.128.IN-ADDR.ARPA. 2000 NS  Q.ISI.EDU.               
                            2000 NS  XX.MIT.EDU.              
                            5000 SOA <SOA information>        
       Q.ISI.EDU.           2000 A   <address of Q.ISI.EDU.>  
       XX.MIT.EDU.          2000 A   <address of XX.MIT.EDU.> 

   SOA serials

      The serial field of the SOA RR for a domain is supposed to be a
      continuously increasing (mod 2**32) value which denotes the

      version of the database.  The idea is that you can tell that a
      zone has changed by comparing serial numbers.  When you change a
      zone, you should increment the serial field of the SOA.

   All RRs with the same name, class, and type should have the same TTL.

      The logic here is that all of them will timeout simultaneously if
      cached and hence the cache can be reliably used.

   Case consistency

      The domain system is supposed to preserve case, but be case
      insensitive.  However, it does nobody any good to put both RRs for
      domain name xxx and XXX in the data base - It merely makes caching
      ambiguous and decreases the efficiency of compression.  This
      consistency should also exist in the duplicate RRs that mark
      delegation in the delegator and delegatee.  For example, if you
      ask the NIC to delegate UZOO.EDU to you, your database shouldn't
      say uzoo.edu.

   Inappropriate use of aliases

      Canonical names are preferred to aliases in all RRs.  One reason
      is that the canonical names are closer to the information
      associated with a name.  A second is that canonical names are
      unique, and aliases are not, and hence comparisons will work.

      In particular, the use of aliases in PTR RRs of the IN-ADDR domain
      or in NS RRs that mark delegation is discouraged.

EXPERIENCES

   This section discusses some unusual situations and common bugs which
   are encountered in the present system, and should be helpful in
   problem determination and tuning.  Put differently, you should try to
   make your code defend against these attacks, and you should expect to
   be the object of complaint if you make these attacks.

   UDP addresses

      When you send a query to a host with multiple addresses, you might
      expect the response to be from the address to which you sent the
      query.  This isn't the case with almost all UNIX implementations.

   UDP checksums

      Many versions of UNIX generate incorrect UDP checksums, and most
      ignore the checksum of incoming UDP datagrams.  The typical
      symptom is that your UNIX domain code works fine with other
      UNIXes, but won't communicate with TOPS-20 or other systems.
      (JEEVES, the TOPS-20 server used for 3 of the 4 root servers,
      ignores datagrams with bad UDP checksums.)

   Making up data

      There are lots of name servers which return RRs for the root
      servers with 99999999 or similar large values in the TTL.  For
      example, some return RRs that suggest that ISIF is a root server.
      (It was months ago, but is no longer.)

      One of the main ideas of the domain system is that everybody can
      get a chunk of the name space to manage as they choose.  However,
      you aren't supposed to lie about other parts of the name space.
      Its OK to remember about other parts of the name space for caching
      or other purposes, but you are supposed to follow the TTL rules.

      Now it may be that you put such records in your server or whatever
      to ensure a server of last resort.  That's fine.  But if you
      export these in answers to queries, you should be shot.  These
      entries get put in caches and never die.

      Suggested domain meta-rule:

         If you must lie, lie only to yourself.

PROBLEM AREAS

   This section discusses some shortcomings in the present system which
   may be addressed in future versions.

   Compression and types

      The present specification attempts to allow name servers and
      resolvers to cache RRs for classes they don't "understand" as well
      as to allow compression of domain names to minimize the size of
      UDP datagrams.  These two goals conflict in the present scheme
      since the only way to expand a compressed name is to know that a
      name is expected in that position.

      One technique for addressing this problem would be to preface
      binary data (i.e. anything but a domain name) with a length octet.

      The high order two bits of the length octet could contain either
      01 or 10, which are illegal for domain names.  To compensate for
      the additional bytes of data, we could omit the RDATA length field
      and terminate each RR with a binary length field of zero.

   Caching non-existent names

      In the present system, a resolver has no standard method for
      caching the result that a name does not exist, which seems to make
      up a larger than expected percentage of queries.  Some resolvers
      create "does not exist" RRs with TTLs to guarantee against
      repetitive queries for a non-existent name.

      A standard technique might be to return the SOA RR for the zone
      (note that only authoritative servers can say name does not exist)
      in the reply, and define the semantics to be that the requester is
      free to assume that the name does not exist for a period equal to
      the MINIMUM field of the SOA.

   Cache conflicts

      When a resolver is processing a reply, it may well decide to cache
      all RRs found in sections of the reply.  The problem is that the
      resolver's cache may already contain a subset of these RRs,
      probably with different TTLs.

      If the RRs are from authoritative data in the answer section, then
      the cache RRs should be replaced.  In other cases, the correct
      strategy isn't completely clear.  Note that if the authoritative
      data's TTL has changed, then the resolver doesn't have enough
      information to make the correct decision in all cases.

      This issue is tricky, and deserves thought.

REFERENCES

   [1]  Mockapetris, P., "Domain Names - Concepts and Facilities",
        RFC-882, USC Information Sciences Institute, November 1983.

   [2]  Mockapetris, P., "Domain Names - Implementation and
        Specification", RFC-883, USC Information Sciences Institute,
        November 1983.

   [3]  Partridge, C., "Mail Routing and the Domain System", RFC-974,
        CSNET-CIC, BBN Laboratories, January 1986.