On Technical Difficulties

It’s great to see the recent discussion of algorithms and archival work that has been prompted by Clifford Lynch’s recent piece for First Monday. Alexis Madrigal’s response in the Atlantic tied together some additional threads and generally did a nice job of situating archival work and the work of history. The record is never perfect, and part of the work of history is learning to work with the records that have survived as a result of care and attention. Both articles also do a nice job of pointing to the significant body of work that has been done on the sociocultural study of algorithms.

But to make his point Madrigal used the Library of Congress Twitter Archive as a case study to illustrate the perils of attempting to archive a large system like Twitter:

Unfortunately, one of the things the library learned was that the Twitter data overwhelmed the technical resources and capacities of the institution. By 2013, the library had to admit that a single search of just the Twitter data from 2006 to 2010 could take 24 hours. Four years later, the archive still is not available to researchers.

As someone who was directly involved with some of the technical work I have a bone to pick with this portrayal. From where I sat (not up on the 6th floor of the Madison Building) the failing at the Library of Congress was not technical in nature but one of policy and governance. Using technical difficulties as an explanation for the failure is simply a convenient excuse for what were actually larger and more interesting problems.

Searching through billions of tweets is by no means a simple matter. But the team I worked on had several plans for how to do it. All of them involved a modest investment in resources (we wanted to use Amazon Web Services at the time) which management was unwilling to invest in. We arrived at these plans after consulting with researchers like Jimmy Lin who had worked at Twitter to set up their own Hadoop infrastructure. The technical infrastructure we had at our disposal at the Library of Congress was only suitable for moving the tweets to a tape archive, where (hopefully) the data still sits.

So there were technical challenges to be sure, but we had a plan to deal with them. The plan was written up and sent up the food chain, and that was the last we heard of it. The policy discussions around why not to invest in the technical resources were, as they say, above my pay grade. But some of them managed to filter down to me. Here were the big ones I heard about through the grape vine:

Can we use Amazon Web Services? At the time the Library of Congress wasn’t using Amazon Web Services — unless you count the World Digital Library Project which was (and still is) a UNESCO project hosted at the Library. Other parts of the Federal Government did use AWS, even in touchy areas like law enforcement — so we knew it could be done. The General Services Administration had a vehicle for using AWS. I think there are lots of good reasons not to use AWS, but the question of why we couldn’t use AWS to provide access to the Twitter data was never satisfactorily answered. At the time I chalked it up to upper management not understanding what it meant to move computing to the cloud, and being unwilling to learn how to adjust their existing mammoth contracts with IBM.

Who is a researcher? The initial donor agreement between Twitter and the Library of Congress indicates that the data should only be made available to bona fide researchers who sign an agreement that they will not use the data for commercial purposes. Access restrictions like this aren’t uncommon for archival collections. But this agreement for the researcher to sign, which was to be agreed upon by the Library of Congress and Twitter, never materialized. Why not?

What is bulk access? The agreement also prohibits bulk access to the data on the web. Is bulk access a thousand tweets, or a million? The agreement doesn’t say. If someone travels to Washington DC, and goes to the Special Collections Reading Room with a few hard drives and wants a copy of the entire archive is that OK? How about a day or an hour? In this hearing before the House Committee on Appropriations Robert Dizard indicated that providing this type of access was possible without additional investment of resources (start listening at 01:16:53, or read the hearing transcript, p. 119). If it was easy and didn’t cost any additional money why didn’t they do this? I bet more than a few researchers would have been overjoyed to have this level of access.

What about deletes? The initial agreement mentions that only data more than 6 months old should be made available. But scuttlebutt while I was working there had it that the initial donor agreement was followed by a series of negotiations about a second agreement to iron out the details. A significant portion of this involved Twitter’s insistence that tweets that were deleted not be distributed by the Library of Congress. This was technically possible because we received a feed of tweets and deletes. But from a policy perspective it proved problematic. What if the person tweeting is a public official, like the President of the United States? Aren’t those tweets part of the public record? On the other hand, people who signed up to Twitter never really agreed to having their tweets saved for the long term. I mean I guess they signed away their ability to disagree. So honoring the rights of people to delete their content is rightly important to Twitter. Enter the conundrum.

So, this is all just the tip of the iceberg, and partly hearsay I suppose, since I wasn’t in the actual rooms where the policy issues were being discussed. If you are interested in reading more Michael Zimmer has a really good article (also for First Monday) that describes the policy issues that the Library of Congress faced around the Twitter Archive. I just wanted to take a brief moment to dig a little deeper into the causes of LC’s failure to provide researcher access to the data — which, last I heard, is still flowing … somewhere.


One more thought I’ll leave you with. If a researcher did have to wait 24 hours to search through billions of tweets to get say, the tweets that used the #blacklivesmatter hashtag, or all the tweets from Wikileaks, or all the tweets from elected representatives in the US Congress, would that be such a bad thing? 24 hours isn’t that long to wait for something you really want to read is it?