Change Details

The current loader git consumes a lot of memory depending on the size of the repository. It's fetching the full packfile of unknown references/refs (filtered by last snapshot's references), then parses the packfile multiple times to load in order contents, directories, revisions, releases and then finishes by creating one snapshot for the visit. References in this context are resolved tip of branches (e.g refs/heads/master, ...) or tags (e.g refs/tags/...). While the memory consumption is not a problem for small (< 200 refs) to medium repositories (<= 500 refs), this can become one on large repositories (> 500), either: 1. The currently unique packfile retrieved at the beginning of the loading is too big (> 4Gib) which fails immediately the ingestion. Nothing gets done. The visit is marked as failed. If that happens too often (thrice consecutively iirc), the origin ends up disabled, so no longer scheduled (up until it's listed again). 2. The ingestion starts but due to concurrency with other loading processes, the ingestion process gets killed. That means partial ingestion of objects got done, but no snapshot nor finalized visit. This is currently a major problem. Current deployment details also implies a heavy disk i/o (which creates ceph problems down the line, which can possibly cascade to outage, see the events we faced during holidays for example...). 3. The last point 2. is also problematic for scheduling further visits for that origin. Nonetheless, if further visit happens somehow, those will skip already ingested objects (which will have still been retrieved again though *without* partial snapshot in between visits). To solve these problems, some work has been investigated and tried. A first naive attempt has been made to iterate over the packfile once and keep a dict of the references (to drop immediately the packfile reference) [1]. This failed as the memory consumption spiked even further. This had the advantage to kill the loading very fast. So, the conclusion of this attempt is that iterating over the packfile multiple times (one iteration for each type of object of our model) is actually not the problem. [1] https://forge.softwareheritage.org/D6377 Another attempt was to modify the loader git to make the ingestion fetch multiple packfiles ([2] [3] with a slight loader-core change required [4])). This has the advantage of naturally taking care of 1. (no more huge packfile). This is done by asking intervals of unknown remote refs, starting by the tags (in natural order) then the branches [5]. The natural order on tags sounds like a proper way to start incrementally load the repository following its history [3]. If we don't follow the history (only [2]), we could fetch first a huge packfile (with mostly everything in it) thus back to square one. This does assume tags exist in the repository (which should mostly be the case). The only limitation seen for that approach is that we now continually discuss with the server to retrieve information (so a time trade-off) during the loading. FWIW, this continuous server discussion approach is what's currently done with the mercurial loader without issues (which is stable now and not as greedy in memory as it used to be, hence one other motivation to align git loader behavior). [2] https://forge.softwareheritage.org/D6386 (ingest in multiple packfile fetch) [3] https://forge.softwareheritage.org/D6392 (follow refs in order) [4] https://forge.softwareheritage.org/D6380. This core loader adaptation actually allows the loader git to create partial visit targeting a snapshot after each packfile consumption. So this makes sure that we create incremental snapshot prior to the final one (if we reach the end of the ingestion). Such adaptation takes care of point 3. (and make subsequent visits do less work even in case of failure). Thanks to @vlorentz which made me realize this, after a couple of nights sleeping on it, it clicked! [5] Another idea (not followed through) would be to ingest some known special references which are assumed highly connected within the graph (e.g."HEAD", "refs/heads/master", "refs/heads/main", "refs/heads/develop", ... others?) at the end of the loading. The hypothesis is that these references are the main connected part of the repository. So starting with those would end up with a huge packfile immediately (so with large repositories at least, we are back to the initial problem). On the contrary, if we start by the other references first, then dealing with the special ones at the end, only a bit more work would be needed to fill in the blanks. That could yet be another optimization which would maybe help if there are no tags in the repository for example. Another consideration we did not follow completely through yet was to use a depth parameter (in the current internal lib used to discuss with the server). It's not completely clear what actual depth number would be a relatively decent and satisfying enough for all repositories out there. It's been slighly tested but dismissed for now due to that question. It's not to be excluded though. It may simply be that this solution composed with the previous points could just be a deeper optimization on reducing the loader's work (especially the part walking the git graph). As another optimization point to further align the git loader with the mercurial loader, it would be interesting to start using the extid table to map what's considered git ids (which will change afaiui) with the revision/release id. Then start using this mapping to actually filter across origins known refs (and not only from the last snapshot of the same origin as currently). That optimization gave a good boost in actually doing less work for ingesting mercurial forks. Filtering early enough known refs (revision/release) would actually reduce further the packfiles (when the extid table is actually filled enough that is). Having described all that, I'm fairly convinced that some approaches described here are a possible way forward (especially the ones already implemented in diffs) which is better than the current status quo. Shall we proceed? Any thoughts, pros or cons arguments welcomed. Cheers,

#+title: Analyze and try to reduce loader-git memory consumption and overall work #+author: vsellier, ardumont The current loader git consumes a lot of memory depending on the size of the repository. It's fetching the full packfile of unknown references/refs (filtered by last snapshot's references), then parses the packfile multiple times to load in order contents, directories, revisions, releases and then finishes by creating one snapshot for the visit. References in this context are resolved tips of branches (e.g refs/heads/master, ...) or tags (e.g refs/tags/...). While the memory consumption is not a problem for small (< 200 refs) to medium repositories (<= 500 refs), this can become one on large repositories (> 500), either: 1. The currently unique packfile retrieved at the beginning of the loading is too big (> 4Gib) which fails immediately the ingestion. Nothing gets done. The visit is marked as failed. If that happens too often (thrice consecutively iirc), the origin ends up disabled, so no longer scheduled (up until it's listed again). 2. The ingestion starts but due to concurrency with other loading processes, the ingestion process gets killed. That means partial ingestion of objects got done, but no snapshot nor finalized visit. This is currently a major problem. Current deployment details also implies a heavy disk i/o (which creates ceph problems down the line, which can possibly cascade to outage, see the events we faced during holidays for example...). 3. The last point 2. is also problematic for scheduling further visits for that origin. Nonetheless, if further visit happens somehow, those will skip already ingested objects (which will have still been retrieved again though *without* partial snapshot in between visits). To solve these problems, some work has been investigated and tried. A first naive attempt has been made to iterate over the packfile once and keep a dict of the references (to drop immediately the packfile reference) [1]. This failed as the memory consumption spiked even further. This had the advantage to kill the loading very fast. So, the conclusion of this attempt is that iterating over the packfile multiple times (one iteration for each type of object of our model) is actually not the problem. [1] https://forge.softwareheritage.org/D6377 Another attempt was to modify the loader git to make the ingestion fetch multiple packfiles ([2] [3] with a slight loader-core change required [4])). This has the advantage of naturally taking care of 1. (no more huge packfile). This is done by asking intervals of unknown remote refs, starting by the tags (in natural order) then the branches [5]. The natural order on tags sounds like a proper way to start incrementally load the repository following its history [3]. If we don't follow the history (only [2]), we could fetch first a huge packfile (with mostly everything in it) thus back to square one. This does assume tags exist in the repository (which should mostly be the case). The only limitation seen for that approach is that we now regularly discuss with the server to retrieve information (so a time trade-off) during the loading. FWIW, this continuous server discussion approach is what's currently done with the mercurial loader without issues (which is stable now and not as greedy in memory as it used to be, hence one other motivation to align git loader behavior). [2] https://forge.softwareheritage.org/D6386 (ingest in multiple packfile fetch) [3] https://forge.softwareheritage.org/D6392 (follow refs tags in order, then branches) [4] https://forge.softwareheritage.org/D6380. This core loader adaptation actually allows the git loader (well DVCS loaders) to create partial visit targeting a snapshot after each internal ingestion loop (so for git after each packfile consumption). So this makes sure that we create incremental snapshot prior to the final one (if we reach the end of the ingestion). Such adaptation takes care of point 3. (and make subsequent visits do less work even in case of failures). Thanks to @vlorentz which made me realize this on that diff's review. After a couple of nights sleeping on it, it clicked! [5] Another idea (not followed through) would be to delay ingestion of some special references which are assumed highly connected within the graph at the end. Typically, e.g. "HEAD", "refs/heads/master", "refs/heads/main", "refs/heads/develop", ... (others?). The hypothesis is that these references are the main connected part of the repository. So starting with those would end up with a huge packfile immediately (so with large repositories at least, we are back to the initial problem). On the contrary, if we start by the other references first, then dealing with the special ones at the end, only a bit more work would be needed to fill in the blanks. That could yet be another optimization which would maybe help if there are no tags in the repository for example. Another consideration we did not follow completely through yet was to use a depth parameter (in the current internal lib used to discuss with the server). It's not completely clear what actual depth number would be a relatively decent and satisfying enough for all repositories out there. It's been slighly tested but dismissed for now due to that question. It's not to be excluded though. It may simply be that this solution composed with the previous points could just be a deeper optimization on reducing the loader's work (especially the part walking the git graph). Not necessarily memory consumption related but still, as another optimization point to further align the git loader with the mercurial loader, it would be interesting to start using the extid table to map what's considered git ids (which will change afaiui) with the revision/release id. Then starting using this mapping would allow filtering even further what we already ingested across origins known refs (and not only from the last snapshot of the same origin as currently). That optimization reduced quite a lot the mercurial loader work especially regarding mercurial forks. In the loader git's case, filtering early enough known refs (revision/release) would actually reduce further the packfiles (providing the extid table is actually filled enough that is). Having described all that, I'm fairly convinced that some approaches described here are a possible way forward (especially the ones already implemented in diffs) which is better than the current status quo. Shall we proceed? Any thoughts, pros or cons arguments welcomed. Cheers,

#+title: Analyze and try to reduce loader-git memory consumption and overall work #+author: vsellier, ardumont The current loader git consumes a lot of memory depending on the size of the repository. It's fetching the full packfile of unknown references/refs (filtered by last snapshot's references), then parses the packfile multiple times to load in order contents, directories, revisions, releases and then finishes by creating one snapshot for the visit. References in this context are resolved tips of branches (e.g refs/heads/master, ...) or tags (e.g refs/tags/...). While the memory consumption is not a problem for small (< 200 refs) to medium repositories (<= 500 refs), this can become one on large repositories (> 500), either: 1. The currently unique packfile retrieved at the beginning of the loading is too big (> 4Gib) which fails immediately the ingestion. Nothing gets done. The visit is marked as failed. If that happens too often (thrice consecutively iirc), the origin ends up disabled, so no longer scheduled (up until it's listed again). 2. The ingestion starts but due to concurrency with other loading processes, the ingestion process gets killed. That means partial ingestion of objects got done, but no snapshot nor finalized visit. This is currently a major problem. Current deployment details also implies a heavy disk i/o (which creates ceph problems down the line, which can possibly cascade to outage, see the events we faced during holidays for example...). 3. The last point 2. is also problematic for scheduling further visits for that origin. Nonetheless, if further visit happens somehow, those will skip already ingested objects (which will have still been retrieved again though *without* partial snapshot in between visits). To solve these problems, some work has been investigated and tried. A first naive attempt has been made to iterate over the packfile once and keep a dict of the references (to drop immediately the packfile reference) [1]. This failed as the memory consumption spiked even further. This had the advantage to kill the loading very fast. So, the conclusion of this attempt is that iterating over the packfile multiple times (one iteration for each type of object of our model) is actually not the problem. [1] https://forge.softwareheritage.org/D6377 Another attempt was to modify the loader git to make the ingestion fetch multiple packfiles ([2] [3] with a slight loader-core change required [4])). This has the advantage of naturally taking care of 1. (no more huge packfile). This is done by asking intervals of unknown remote refs, starting by the tags (in natural order) then the branches [5]. The natural order on tags sounds like a proper way to start incrementally load the repository following its history [3]. If we don't follow the history (only [2]), we could fetch first a huge packfile (with mostly everything in it) thus back to square one. This does assume tags exist in the repository (which should mostly be the case). The only limitation seen for that approach is that we now continualregularly discuss with with the server to retrieve information (so a time trade-off) during the loading. FWIW, this continuous server discussion approach is what's currently done with the mercurial loader without issues (which is stable now and not as greedy in memory as it used to be, hence one other motivation to align git loader behavior). [2] https://forge.softwareheritage.org/D6386 (ingest in multiple packfile fetch) [3] https://forge.softwareheritage.org/D6392 (follow refs tags in orderer, then branches) [4] https://forge.softwareheritage.org/D6380. This core loader adaptation actually allows the git loader git(well DVCS loaders) to create partial visit targeting a snapshot after each packfile after each internal ingestion loop (so for git after each packfile consumption). So this makes sure that we create incremental snapshot prior to the final one (if we reach the end of the ingestion). Such adaptation takes care of point 3. (andmakes sure that we create incremental snapshot prior to the final one (if we reach the make subsequent visits do less work even in casend of the ingestion). Such adaptation takes care of failure)point 3. Thanks to @vlorentz which(and make subsequent visits do less work even in case of failures). Thanks to @vlorentz which made me realize this, this on that diff's review. aAfter a couple of nights sleeping on it, it clicked! [5] Another idea (not followed through) would be to delay ingestion of some known special referencesspecial references which are assumed highly connected within the graph at the end. Typically, which are assumed highly connected within the graph (e.g. "HEAD", "refs/heads/master", "refs/heads/main", "refs/heads/develop", ... "refs/heads/main", "refs/heads/develop", ... (others?) at the end of the loading. The hypothesis is that these references are the main connected part of the hypothesis is that these references are the main connected part of the repository. So starting with those would end up with a huge packfile immediately (so starting with those would end up with a huge packfile immediately (so with largewith large repositories at least, we are back to the initial problem). On the contrary, repositories at least, we are back toif we start by the initial problem). On the contraryother references first, if we startthen dealing with the special ones at the by the other references firstend, then dealing with the special ones atonly a bit more work would be needed to fill in the end,blanks. only a bitThat could yet be more work would be needed to fill in the blanks. That could yet be another optimizationanother optimization which would maybe help if there are no tags in the repository for which would maybe help if there are no tags in the repository for example. Another consideration we did not follow completely through yet was to use a depth parameter (in the current internal lib used to discuss with the server). It's not completely clear what actual depth number would be a relatively decent and satisfying enough for all repositories out there. It's been slighly tested but dismissed for now due to that question. It's not to be excluded though. It may simply be that this solution composed with the previous points could just be a deeper optimization on reducing the loader's work (especially the part walking the git graph). ANot necessarily memory consumption related but still, as another optimization point to further align the git loader with the mercurial loader, it would be interesting to start it would be interesting to start using the extid table to map what's considered git ids (which will change afaiui) with (which will change afaiui) with the revision/release id. Then starting using this mappinging would allow filtering even to actually filterfurther what we already ingested across origins known refs (and not only from the last snapshot of the snapshot of the same origin as currently). That optimization gave a good boost in actually doing lessreduced quite a lot the mercurial loader work for ingestespecially regarding mercurial forks. Filtering early enough known refs (revision/release)In the loader git's case, would actually reduce further the packfiles (when the extid table is actually filledfiltering early enough known refs (revision/release) would actually reduce further the packfiles (providing the extid table is actually filled enough that is). Having described all that, I'm fairly convinced that some approaches described here are a possible way forward (especially the ones already implemented in diffs) which is better than the current status quo. Shall we proceed? Any thoughts, pros or cons arguments welcomed. Cheers,