What to put in a binary data file's header

问题:

I have a simulation that reads large binary data files that we create (10s to 100s of GB). We use binary for speed reasons. These files are system dependent, converted from text files on each system that we run, so I'm not concerned about portability. The files currently are many instances of a POD struct, written with fwrite.

I need to change the struct, so I want to add a header that has a file version number in it, which will be incremented anytime the struct changes. Since I'm doing this, I want to add some other information as well. I'm thinking of the size of the struct, byte order, and maybe the svn version number of the code that created the binary file. Is there anything else that would be useful to add?

回答1:

In my experience, second-guessing the data you'll need is invariably wasted time. What's important is to structure your metadata in a way that is extensible. For XML files, that's straightforward, but binary files require a bit more thought.

I tend to store metadata in a structure at the END of the file, not the beginning. This has two advantages:

• Truncated/unterminated files are easily detected.
• Metadata footers can often be appended to existing files without impacting their reading code.

The simplest metadata footer I use looks something like this:

struct MetadataFooter{
  char[40] creatorVersion;
  char[40] creatorApplication;
  .. or whatever
} 

struct FileFooter
{
  int64 metadataFooterSize;  // = sizeof(MetadataFooter)
  char[10] magicString;   // a unique identifier for the format: maybe "MYFILEFMT"
};

After the raw data, the metadata footer and THEN the file footer are written.

When reading the file, seek to the end - sizeof(FileFooter). Read the footer, and verify the magicString. Then, seek back according to metadataFooterSize and read the metadata. Depending on the footer size contained in the file, you can use default values for missing fields.

As KeithB points out, you could even use this technique to store the metadata as an XML string, giving the advantages of both totally extensible metadata, with the compactness and speed of binary data.

回答2:

For large binaries I'd look seriously at HDF5 (Google for it). Even if it's not something you want to adopt it might point you in some useful directions in designing your own formats.

回答3:

For large binaries, in addition to the version number I tend to put a record count and CRC, the reason being that large binaries are much more prone to get truncated and/or corrupted over time or during transfer than smaller ones. I found recently to my horror that Windows does not handle this well at all, as I used explorer to copy about 2TB across a couple of hundred files to an attached NAS device, and found 2-3 files on each copy were damaged (not completely copied).

回答4:

An identifier for the type of the file would be useful if you will have other structures written to binary files later on. Maybe this could be a short string so you can see by a look into the file (via hex editor) what it contains.

回答5:

If they're that large, I'd reserve a healthy chunk (64K?) of space at the beginning of the file and put the metadata there in XML format followed by an end-of-file character (Ctrl-Z for DOS/Windows, ctrl-D for unix?). That way you can examine and parse the metadata easily with the wide range of toolsets out there for XML.

Otherwise I go with what other people have already said: timestamp for file creation, identifier for which machine it's created on, basically anything else that you can think of for diagnostic purposes. And ideally you would include the definition of the structure format itself. If you are changing the structure often, it's a big pain to maintain the proper version of code around to read various formats of old datafiles.

One big advantage of HDF5 as @highpercomp has mentioned, is that you just don't need to worry about changes in the structure format, as long as you have some convention of what the names and datatypes are. The structure names and datatypes are all stored in the file itself, so you can blow your C code to smithereens and it doesn't matter, you can still retrieve data from an HDF5 file. It lets you worry less about the format of data and more on the structure of data, i.e. I don't care about the sequence of bytes, that's HDF5's problem, but I do care about field names and the like.

Another reason I like HDF5 is you can choose to use compression, which takes a very small amount of time and can give you huge wins in storage space if the data is slowly-changing or mostly the same except for a few errant blips of interestingness.

回答6:

@rstevens said 'an identifier for the type of file'...sound advice. Conventionally, that's called a magic number and, in a file, isn't a term of abuse (unlike in code, where it is a term of abuse). Basically, it is some number - typically at least 4 bytes, and I usually ensure that at least one of those bytes is not ASCII - that you can use to validate that the file is of the type you expect with a low probability of being confused. You can also write a rule in /etc/magic (or local equivalent) to report that files containing your magic number are your special file type.

You should include a file format version number. However, I would recommend not using the SVN number of the code. Your code may change when the file format does not.

回答7:

In addition to whatever information you need for schema versioning, add details that may be of value if you are troubleshooting an issue. For example:

• timestamps of when the file was created and update (if applicable).
• the version string from the build (ideally you have a version string that is auto-incremented on every 'official' build ... this is different to the file schema version).
• the name of the system creating the file, and maybe other statistics that are relevant to your app

We find this is very useful (a) in getting information we would otherwise have to ask the customer to provide and (b) getting correct information -- it is amazing how many customers report they are running a different version of the software to what the data claims!

回答8:

You might consider putting a file offset in a fixed position in the header, which tells you where the actual data begins in the file. This would let you change the size of the header when needed.

In a couple of cases, I put the value 0x12345678 into the header so I could detect if the file format, matched the endianism of the machine that was processing it.

回答9:

As my experience with telecom equipment configuration and firmware upgrades shows you only really need several predefined bytes at the begin (this is important) which starts from version (fixed part of header). Rest of header is optional, by indicating proper version you can always show how to process it. Important thing here is you'd better place 'variable' part of header at the end of file. If you plan operations on header without modifying file content itself. Also this simplify 'append' operations which should recalculate variable header part.

Nice to have features for fixed size header (at the begin):

• Common 'length' field (including header).
• Something like CRC32 (including header).

OK, for variable part XML or some pretty extensible format in header is good idea but is it really needed? I had lot of experience with ASN encoding... in most cases its usage was overshot.

Well, maybe you will have additional understanding when you look at things like TPKT format which is described in RFC 2126 (chapter 4.3).

回答10:

If you are putting a version number in the header you can change that version anytime you need to change the POD struct or add new fields to the header.

So don't add stuff to the header now because it might be interesting. You are just creating code that you have to maintain but that has little real value.

回答11:

For large files, you might want to add data definitions, so your file format becomes self-describing.

回答12:

My variation combines Roddy and Jason S's approaches.

In summary - put formatted text metadata at the end of the file with a way to determine its length stored elsewhere.

1) Put an length field at the beginning of your file so you know the length of the metadata at the end rather than assuming a fixed length. That way, to get the metadata you just read that fixed-length initial field and then get the metadata blob from the end of file.

2) Use XML or YAML or JSON for the metadata. This is especially useful/safe if the metadata is appended at the end because nobody reading the file is going to automatically think it's all XML just because it starts with XML.

The only disadvantage in this approach is when your metadata grows, you have to update both the head of the file and the tail but it's likely other parts will have been updated anyway. If it's just updating trivia like a last-accessed date then the metadata length won't change so it only needs an update in-place.