6 The target audience for this document is developers who wish to understand
7 the new lease database (leasedb) planned to be added in Tahoe-LAFS v1.11.0.
13 A "lease" is a request by an account that a share not be deleted before a
14 specified time. Each storage server stores leases in order to know which
15 shares to spare from garbage collection.
20 The leasedb will replace the current design in which leases are stored in
21 the storage server's share container files. That design has several
24 - Updating a lease requires modifying a share container file (even for
25 immutable shares). This complicates the implementation of share classes.
26 The mixing of share contents and lease data in share files also led to a
27 security bug (ticket `#1528`_).
29 - When only the disk backend is supported, it is possible to read and
30 update leases synchronously because the share files are stored locally
31 to the storage server. For the cloud backend, accessing share files
32 requires an HTTP request, and so must be asynchronous. Accepting this
33 asynchrony for lease queries would be both inefficient and complex.
34 Moving lease information out of shares and into a local database allows
35 lease queries to stay synchronous.
37 Also, the current cryptographic protocol for renewing and cancelling leases
38 (based on shared secrets derived from secure hash functions) is complex,
39 and the cancellation part was never used.
41 The leasedb solves the first two problems by storing the lease information in
42 a local database instead of in the share container files. The share data
43 itself is still held in the share container file.
45 At the same time as implementing leasedb, we devised a simpler protocol for
46 allocating and cancelling leases: a client can use a public key digital
47 signature to authenticate access to a foolscap object representing the
48 authority of an account. This protocol is not yet implemented; at the time
49 of writing, only an "anonymous" account is supported.
51 The leasedb also provides an efficient way to get summarized information,
52 such as total space usage of shares leased by an account, for accounting
55 .. _`#1528`: https://tahoe-lafs.org/trac/tahoe-lafs/ticket/1528
61 A share is stored as a collection of objects. The persistent storage may be
62 remote from the server (for example, cloud storage).
64 Writing to the persistent store objects is in general not an atomic
65 operation. So the leasedb also keeps track of which shares are in an
66 inconsistent state because they have been partly written. (This may
67 change in future when we implement a protocol to improve atomicity of
68 updates to mutable shares.)
70 Leases are no longer stored in shares. The same share format is used as
71 before, but the lease slots are ignored, and are cleared when rewriting a
72 mutable share. The new design also does not use lease renewal or cancel
73 secrets. (They are accepted as parameters in the storage protocol interfaces
74 for backward compatibility, but are ignored. Cancel secrets were already
75 ignored due to the fix for `#1528`_.)
77 The new design needs to be fail-safe in the sense that if the lease database
78 is lost or corruption is detected, no share data will be lost (even though
79 the metadata about leases held by particular accounts has been lost).
85 A "crawler" is a long-running process that visits share container files at a
86 slow rate, so as not to overload the server by trying to visit all share
87 container files one after another immediately.
89 The accounting crawler replaces the previous "lease crawler". It examines
90 each share container file and compares it with the state of the leasedb, and
91 may update the state of the share and/or the leasedb.
93 The accounting crawler may perform the following functions (but see ticket
94 #1834 for a proposal to reduce the scope of its responsibility):
96 - Remove leases that are past their expiration time. (Currently, this is
97 done automatically before deleting shares, but we plan to allow expiration
98 to be performed separately for individual accounts in future.)
100 - Delete the objects containing unleased shares — that is, shares that have
101 stable entries in the leasedb but no current leases (see below for the
102 definition of "stable" entries).
104 - Discover shares that have been manually added to storage, via ``scp`` or
105 some other out-of-band means.
107 - Discover shares that are present when a storage server is upgraded to
108 a leasedb-supporting version from a previous version, and give them
111 - Recover from a situation where the leasedb is lost or detectably
112 corrupted. This is handled in the same way as upgrading from a previous
115 - Detect shares that have unexpectedly disappeared from storage. The
116 disappearance of a share is logged, and its entry and leases are removed
123 An account holds leases for some subset of shares stored by a server. The
124 leasedb schema can handle many distinct accounts, but for the time being we
125 create only two accounts: an anonymous account and a starter account. The
126 starter account is used for leases on shares discovered by the accounting
127 crawler; the anonymous account is used for all other leases.
129 The leasedb has at most one lease entry per account per (storage_index,
130 shnum) pair. This entry stores the times when the lease was last renewed and
131 when it is set to expire (if the expiration policy does not force it to
132 expire earlier), represented as Unix UTC-seconds-since-epoch timestamps.
134 For more on expiration policy, see `docs/garbage-collection.rst
135 <../garbage-collection.rst>`__.
141 The leasedb holds an explicit indicator of the state of each share.
143 The diagram and descriptions below give the possible values of the "state"
144 indicator, what that value means, and transitions between states, for any
145 (storage_index, shnum) pair on each server::
148 # STATE_STABLE -------.
151 # STATE_COMING | | STATE_GOING
154 # '----- NONE <------'
157 **NONE**: There is no entry in the ``shares`` table for this (storage_index,
158 shnum) in this server's leasedb. This is the initial state.
160 **STATE_COMING**: The share is being created or (if a mutable share)
161 updated. The store objects may have been at least partially written, but
162 the storage server doesn't have confirmation that they have all been
165 **STATE_STABLE**: The store objects have been completely written and are
166 not in the process of being modified or deleted by the storage server. (It
167 could have been modified or deleted behind the back of the storage server,
168 but if it has, the server has not noticed that yet.) The share may or may not
171 **STATE_GOING**: The share is being deleted.
176 • **STATE_GOING** → **NONE**
178 trigger: The storage server gains confidence that all store objects for
179 the share have been removed.
183 1. Remove the entry in the leasedb.
185 • **STATE_STABLE** → **NONE**
187 trigger: The accounting crawler noticed that all the store objects for
192 1. Remove the entry in the leasedb.
194 • **NONE** → **STATE_COMING**
196 triggers: A new share is being created, as explicitly signalled by a
197 client invoking a creation command, *or* the accounting crawler discovers
202 1. Add an entry to the leasedb with **STATE_COMING**.
204 2. (In case of explicit creation) begin writing the store objects to hold
207 • **STATE_STABLE** → **STATE_COMING**
209 trigger: A mutable share is being modified, as explicitly signalled by a
210 client invoking a modification command.
214 1. Add an entry to the leasedb with **STATE_COMING**.
216 2. Begin updating the store objects.
218 • **STATE_COMING** → **STATE_STABLE**
220 trigger: All store objects have been written.
224 1. Change the state value of this entry in the leasedb from
225 **STATE_COMING** to **STATE_STABLE**.
227 • **NONE** → **STATE_STABLE**
229 trigger: The accounting crawler discovers a complete share.
233 1. Add an entry to the leasedb with **STATE_STABLE**.
235 • **STATE_STABLE** → **STATE_GOING**
237 trigger: The share should be deleted because it is unleased.
241 1. Change the state value of this entry in the leasedb from
242 **STATE_STABLE** to **STATE_GOING**.
244 2. Initiate removal of the store objects.
247 The following constraints are needed to avoid race conditions:
249 - While a share is being deleted (entry in **STATE_GOING**), we do not accept
250 any requests to recreate it. That would result in add and delete requests
251 for store objects being sent concurrently, with undefined results.
253 - While a share is being added or modified (entry in **STATE_COMING**), we
256 - Creation or modification requests for a given mutable share are serialized.
259 Unresolved design issues
260 ------------------------
262 - What happens if a write to store objects for a new share fails
263 permanently? If we delete the share entry, then the accounting crawler
264 will eventually get to those store objects and see that their lengths
265 are inconsistent with the length in the container header. This will cause
266 the share to be treated as corrupted. Should we instead attempt to
267 delete those objects immediately? If so, do we need a direct
268 **STATE_COMING** → **STATE_GOING** transition to handle this case?
270 - What happens if only some store objects for a share disappear
271 unexpectedly? This case is similar to only some objects having been
272 written when we get an unrecoverable error during creation of a share, but
273 perhaps we want to treat it differently in order to preserve information
274 about the storage service having lost data.
276 - Does the leasedb need to track corrupted shares?
282 Clients will have key pairs identifying accounts, and will be able to add
283 leases for a specific account. Various space usage policies can be defined.
285 Better migration tools ('tahoe storage export'?) will create export files
286 that include both the share data and the lease data, and then an import tool
287 will both put the share in the right place and update the recipient node's